IJARCCE adheres to the suggestive parameters outlined by the University Grants Commission (UGC) for peer-reviewed journals, upholding high standards of research quality, ethical publishing, and academic excellence.
Intelligent Stock Price Prediction and Investment Recommendation System using Machine Learning and Deep LSTM Framework for Equity Price Forecasting with Real-Time Volatility Shock Detection
N. Abilasha, S. Hariharan, Dr. K. Akila*
Abstract: Forecasting equity prices remains difficult because financial markets are inherently non-stationary, heavily influenced by sentiment, and prone to abrupt volatility spikes. Standard machine learning pipelines usually forecast point estimates without factoring in sudden tail-risk events, causing models to perform poorly during major market disruptions. In this work, we implement a practical dual-model prediction framework that pairs a two-layer PyTorch Long Short-Term Memory (LSTM) network with an ensemble Random Forest regressor. To tackle the problem of price instability, we introduce a dedicated volatility detection module that monitors rolling 20-day return Z-scores, intraday trading spreads, and dynamic Bollinger Band boundaries at both 2-sigma and 3-sigma thresholds. Instead of predicting raw unnormalized prices—which suffer from long-term scale drift—our LSTM models the relative price ratio between consecutive trading sessions across a 10-day lookback window. We tested the pipeline on six diverse equities across multiple exchanges: Tesla, Microsoft, META, Accenture, Samsung Electronics, and ICICI Bank, totaling more than 25,000 historical trading sessions. Our experiments show that the models achieve strong goodness- of-fit, reaching R² values of 0.9984 for Microsoft and 0.9862 for Tesla, with root mean squared errors of 4.95 USD and 10.78 USD, respectively. Furthermore, our volatility corridor correctly flagged 98.4% of historical price shocks, allowing the built-in recommendation engine to adjust stop-loss limits and protect capital. These results confirm that combining sequential deep learning with statistical shock boundaries offers a dependable foundation for real-world trading decision support.
Keywords: Equity Forecasting, Long Short-Term Memory, Random Forest, Bollinger Bands, Volatility Shocks, Stop- Loss Optimization, Financial Time Series.
AI-Powered Storage System Classifier: Intelligent File Tiering Using Access Patterns
Nooka Harshavardhan, T. V. Sai Prudhvi
Abstract: As digital systems generate and retain increasing volumes of data, selecting an appropriate storage tier becomes important for balancing access speed, storage cost, and resource utilization. This paper presents an AI-powered storage system classifier that categorizes files into Hot, Warm, and Cold storage tiers based on observed file metadata and access behaviour. The proposed system collects metadata such as access frequency, time since last access, file age, file size, and recent access trends. These features are supplied to a Random Forest classification model, which predicts the most suitable storage tier. Frequently accessed files are assigned to Hot storage, files accessed occasionally are assigned to Warm storage, and rarely accessed files are assigned to Cold storage. The design includes a monitoring process that periodically updates file activity and supports reclassification when usage patterns change. Python, Pandas, NumPy, scikit-learn, Streamlit, and SQLite are proposed for the implementation, with a local filesystem or cloud object storage such as AWS S3 serving as the storage backend. This paper describes the system design and evaluation plan; measured accuracy and performance results are not claimed until implementation and testing are completed.
Keywords: storage tiering, machine learning, Random Forest, file classification, access patterns, Hot storage, Warm storage, Cold storage, intelligent storage management
AI-Based Resume Filter System: Intelligent Candidate Screening and Job Matching
Nooka Harshavardhan, T. V. Sai Prudhvi, R. Bharath
Abstract: Recruitment teams often receive a large number of resumes for a single job opening. Manually reviewing every application can consume significant time and may lead to inconsistent shortlisting. This paper proposes an AI- based Resume Filter System that assists recruiters by extracting information from resumes, identifying relevant skills, comparing candidate profiles with a job description, and producing an explainable shortlist for human review. The proposed workflow accepts resumes in PDF or DOCX format, extracts and preprocesses text, identifies education, experience, skills, and relevant keywords, and represents candidate information for comparison with job requirements. A matching and ranking component can combine skill coverage, relevant experience, education requirements, and semantic similarity to generate a candidate relevance score. The system is intended as a decision-support tool rather than an autonomous hiring authority. Recruiters retain responsibility for reviewing candidates and making selection decisions. The paper describes the proposed architecture, methodology, technology options, evaluation plan, limitations, and safeguards related to privacy, bias, and transparency. No measured accuracy or recruitment improvement is claimed until the system is implemented and evaluated on an appropriate dataset.
SkillSwap: A Secure Peer-to-Peer Platform for Non-Monetary Skill Exchange and Accountable Collaborative Learning in Higher Education
Kinjal Dnyaneshwar Mali, Prof. Manoj V. Nikum, Dr. D. M. Patel
Abstract: University students face economic and pedagogical barriers when acquiring practical technical skills: commercial platforms and tutoring services require recurring fees, massive open online courses report low completion, and informal messaging groups lack structure, scheduling and accountability. This paper presents SkillSwap, a web- based peer-to-peer platform on which students exchange competencies they can teach for competencies they wish to learn, without monetary transactions. The system follows a three-tier architecture comprising a vanilla JavaScript client, a Python Flask 3.0 application tier with 21 REST routes, and a Supabase-hosted PostgreSQL 15 persistence tier that enforces Row Level Security (RLS), JSON Web Token (JWT) authentication and trigger-based profile provisioning. A bilateral workflow governs peer discovery through debounced live autocomplete, exchange negotiation, session scheduling, rescheduling and dual-confirmation completion, while every state change is written to an audit table. Validation comprised 28 functional test cases and benchmarks of six transactional endpoints under one to 200 concurrent simulated users. All 28 test cases passed. The live autocomplete endpoint recorded a median latency of 28 ms, every endpoint met its 95th-percentile service target, and the platform sustained 890 requests per second at 200 concurrent users with a 0.00% error rate. The results indicate that a structured, secure and auditable barter model for peer learning is technically feasible on commodity infrastructure.
Real-Time Smartphone-Based Driver Drowsiness Detection Using Multi-Cue Facial Landmark Fusion and Graded Alerting
Lalitha P.C., Rajendhran A., Deepa T. and Akila K.*
Abstract: A drowsy driver reacts slowly, judges poorly, and often gives no outward sign of it. Most systems built to catch this rely on infrared cameras, wearable sensors, or heavy deep-learning hardware, so they rarely reach ordinary drivers. This paper describes a drowsiness detection and alert system that runs completely on an Android smartphone and needs nothing but its front camera. Every sampled video frame goes through MediaPipe face detection and landmark extraction. From the landmarks we measure four cues: eye closure and blinking (Eye Aspect Ratio, EAR), yawning (Mouth Aspect Ratio, MAR), and head-pose deviation. The cues are collected over a sliding window and merged into one weighted fatigue score between 0 and 100%. A threshold-based decision engine turns that score into one of four states: Normal, Warning, Drowsy, or Severe Drowsiness. Each state has its own response, which grows from an on-screen warning to audio and vibration alarms and, in severe cases, an optional emergency notification. Sessions, events and alerts are saved in a local SQLite database that feeds an offline dashboard, with optional cloud sync. The app is built with Flutter, OpenCV, MediaPipe and TensorFlow Lite. We explain the architecture, the scoring and decision rules, and an evaluation protocol based on accuracy, precision, recall, response time and false-alert rate, and we report eight functional test cases. The results indicate that a cheap, private, graded monitoring pipeline is practical on ordinary phones.
AI-BASED FILE ACCESS PREDICTION SYSTEM FOR SECURE INFORMATION STORAGE
Adithya Pratheep, Vidya Sree DV, Dr. P. Durgadevi
Abstract: Every time someone opens, edits or downloads a file, a record is written somewhere. Over weeks and months those records pile up far beyond what an administrator can read line by line, and odd behaviour is easy to miss in all that noise. This paper describes a prototype that looks at the context around each access event and labels it as normal or suspicious. The inputs include the time of access, IP address, location, how often the user requests files, how sensitive the file is, and what the user has done before. Once the records have been checked and cleaned, they go to a Random Forest classifier. A risk and alert layer acts on its output, and an audit store keeps every event with its prediction so it can be reviewed later. A Streamlit dashboard brings activity, predictions, risk levels and alerts together in one place. The system sits beside existing authentication and authorization controls and does not try to replace them. We present it as a working prototype and a methodology, not as a finished product. We report no accuracy figures, because the model has not yet been tested on an independently prepared dataset. The sections that follow cover the motivation, related work, architecture, feature preparation, prediction workflow, testing plan and directions for future work.
Keywords: file-access prediction, information storage security, Random Forest, anomaly detection, user behaviour analysis, audit logging.
Abstract: Hostel residents frequently face maintenance-related problems such as electrical faults, water leakage, plumbing issues, and cleaning requirements. Traditional complaint reporting methods often depend on verbal communication or manual registers, which can make complaint tracking, prioritization, and status monitoring difficult. This paper presents the Smart Hostel Issue Reporter, a web-based complaint management system designed to provide a structured and centralized platform for reporting and managing hostel complaints. The proposed system provides a student interface for submitting complaints and monitoring their progress, along with an administrator interface for reviewing, assigning, and updating complaints. The application uses React.js for the frontend, Node.js and Express.js for backend services, and MongoDB for data storage. Authentication and role-based access are used to separate student and administrator operations. In addition to web-based reporting, an ESP8266-based physical reporting interface is considered as a supplementary reporting mechanism using an OLED display, push button, buzzer, and indicator LEDs. The system aims to improve complaint visibility, record maintenance, and communication between hostel residents and administrators. The hardware interface is treated as a prototype, and complete hardware-to-backend integration remains a part of further development.
Keywords: Hostel Management, Complaint Management System, Web Application, React.js, Node.js, MongoDB, ESP8266, Internet of Things (IoT).
Smart Parcel Receiving System: A Review of IoT-Based Secure Unattended Delivery
A. Yashwitha, Amrutha K, Ashwini N. R., K. Bindu, Jayasubha J.
Abstract: This structured narrative review examines IoT-based parcel receiving boxes and lockers, focusing on platforms, parcel detection, access authentication, notifications, and evaluation. Thirteen sources are discussed and eight compared. One study reports measured timings and another uses simulation, limiting cross-system comparison. The review identifies gaps in quantitative evaluation, order-system integration, credential separation, connectivity, and security analysis. The related project has implemented its software layer; physical hardware integration and performance evaluation remain planned.
Keywords: Internet of Things, Last-mile delivery, One-time password, Smart parcel box, Unattended delivery.
Croplytics: A Python-Based Smart Agriculture Assistant for Leaf Image Handling, Crop-Yield Estimation and Carbon-Footprint Estimation
Miss. Hemani Pravinsing Girase, Prof. Manoj V. Nikum
Abstract: Croplytics is a Python-based smart agriculture web application that gathers a few farming-support utilities on a single platform: plant-leaf image handling, crop-yield estimation, carbon-footprint estimation and agricultural product information. The backend is written in Python with the Flask framework, which receives browser requests, passes them to the relevant application logic and returns the result. Pillow and NumPy handle image operations, Pandas organizes agricultural datasets, and scikit-learn supports the yield-estimation workflow whenever a trained and compatible model is available. SQLite stores user and history records, and the interface is built with HTML, CSS and JavaScript. The application also provides user registration and login, with validation on both the browser and the server, salted password hashing and parameterized database queries. Croplytics is an educational prototype. Its image-analysis and prediction outputs should be read as demonstrations or estimates until the models and calculations behind them have been validated independently. This paper presents the motivation, background, analysis and design, module-wise implementation, validation approach, observed interface and possible future work.
SKILLPATH AI: NEXT-GENERATION ADAPTIVE AI LEARNING ROADMAP PLATFORM
Bhavika Suresh Vadile, Mr. Manoj V. Nikum
Abstract: Self-directed learners in computer science and software engineering often face fragmented learning resources, rigid course schedules and limited personalized guidance. This paper presents SkillPath AI, an adaptive learning roadmap platform that converts a learner’s target skill, proficiency level, available study time and career goals into a structured day-by-day learning plan. Powered by Groq’s Tensor Streaming Processor (TSP) and Meta’s Llama- 3.3 70B Versatile model, the platform reports curriculum-generation latency below 2.5 seconds. Its modules include an algorithmic curriculum architect, an interactive daily focus tracker with learning streaks, a context-aware AI code tutor and a scenario-based diagnostic quiz runner. A dual-engine persistence layer uses Supabase PostgreSQL when cloud credentials are available and falls back to local SQLite3 otherwise. The project report records 24 functional test cases with a 100% pass rate, a 90.4% reduction in curriculum-generation latency and 5.1× higher token throughput compared with conventional cloud LLM endpoints. These figures represent the results reported for the implemented project.
Keywords: Adaptive Learning, Intelligent Tutoring Systems, Large Language Models, Groq TSP, Llama-3.3 70B, Curriculum Generation, AI Code Tutor.
ApnaRoute: Smart Route Finder and Travel Cost Calculator
Vrushali Rajput, Prof.Manoj V. Nikum
Abstract: ApnaRoute is a web-based route-planning application designed to bring route search, distance and travel- time estimation, fuel-cost calculation and nearby-service discovery into a single interface. Users enter an origin and destination, which are geocoded using Photon with Nominatim as a fallback, and view the locations and route on an interactive OpenStreetMap-based map. The application can present up to three route alternatives—Fastest, Balanced and Scenic—with distance, estimated duration and fuel-cost information. Fuel requirements and estimated cost are calculated from the route distance, user-entered vehicle mileage and fuel price. The system also supports live device location, nearby petrol pumps, hospitals and restaurants, and private saved routes for registered users. The implementation uses HTML, CSS, Bootstrap and JavaScript, Python Flask, Leaflet, OpenStreetMap services and Supabase PostgreSQL with Row Level Security. A sample route test from Kadamwadi, Kolhapur, to Baramati returned 219.3 km and an estimated travel time of 3 hours 7 minutes for the Fastest route; the reported fuel estimate was ■1,412 under the calculator inputs used during that test. The project demonstrates an integrated, open- data-based travel-planning prototype. Travel time is an estimate and does not include live traffic.
Abstract: Maintaining personal fitness requires consistent attention to body measurements, exercise, food intake, and activity history. When these activities are managed through separate applications or manual records, users may find it difficult to maintain an organized view of their progress. FitZen – Personal Fitness and Wellness Tracker is a web- based application designed to bring these everyday fitness functions together within a single platform. The system provides user registration and profile management, a Body Mass Index (BMI) calculator, goal-oriented workout suggestions, general diet recommendations, daily calorie logging, and fitness progress records. The project uses Python Flask for backend routing and application logic, HTML5, CSS3, JavaScript and Bootstrap for the responsive user interface, and Supabase PostgreSQL for relational data storage. Database Row Level Security (RLS), password hashing, session checks, and environment-variable based credential management are incorporated to support user data privacy. The project report describes unit, integration, system, and security testing of the main workflows. FitZen is intended as a beginner-friendly wellness support tool; its BMI output and recommendations are informational and are not substitutes for professional medical advice.
Dr. Beenarani Manoj, Deepak G, Praveen K, Varun VV
Abstract: Choosing a movie from a large catalogue can be difficult because viewers must explore titles, genres, descriptions and other details before finding films that match their interests. This paper presents a content-based Movie Recommendation System that analyses movie metadata and recommends films with characteristics similar to a user- selected title. The system considers attributes such as genre, keywords, overview, cast, director, rating and release year. Textual metadata is transformed into numerical vectors using Term Frequency–Inverse Document Frequency (TF-IDF), and Cosine Similarity is used to compare movie representations. The resulting similarity scores are used to rank candidate films and present relevant recommendations. The project also provides a user-facing discovery interface with title-based search, genre and rating filters, and a locally maintained watchlist. The current implementation focuses on content-based recommendation using TF-IDF and Cosine Similarity. Collaborative filtering, hybrid recommendation, deep-learning embeddings, review sentiment analysis, continuous learning and a mobile client are proposed as future enhancements. The project presentation identifies Precision@K, Recall@K, relevance, coverage, diversity and response time as evaluation measures, but leaves numerical results pending; therefore, no unverified performance values are claimed in this paper.
Keywords: Movie Recommendation, Content-Based Filtering, TF-IDF, Cosine Similarity, Machine Learning, Movie Metadata, Information Retrieval.
AI Resume Analyzer Using RoBERTa for Semantic Candidate-Job Matching
Mr.Vijay Dilip Wagh, Prof. Manoj V. Nikum, Prof. Kaustubh C. Bhave
Abstract: Screening resumes by hand is slow, and the judgement applied differs from one recruiter to another. Applicant Tracking Systems that depend on keyword overlap reduce the workload but overlook candidates whose resumes describe the same skill in different words. This paper presents an AI Resume Analyzer that uses RoBERTa, a robustly optimised BERT-style transformer, to read resumes in context and match them with job requirements. Natural language processing techniques extract skills, qualifications, experience, education, certifications and projects from uploaded resumes, and a fine-tuned RoBERTa model estimates how closely each resume fits a given job. The system returns a relevance score, a suitability category and lists of matched and missing skills, which are shown to the recruiter on a web dashboard. The architecture separates a Java Spring Boot service, which manages authentication, storage and the interface, from a Python engine built on Hugging Face Transformers, PyTorch and spaCy. An evaluation plan against a TF-IDF keyword baseline, using accuracy, precision, recall, F1-score, Precision@k and NDCG, is also described. The proposed system aims to make first-round screening faster, more consistent and easier to explain.
Keywords: AI Resume Analyzer, RoBERTa, Natural Language Processing, Transformer, Candidate-Job Matching, Applicant Tracking System, Skill Extraction, Explainable Screening
Miss. Krutika Pravin Borse, Prof. Manoj V. Nikum, Prof. Mrs. S. N. Raul
Abstract: Academic research papers are long and densely structured, and readers, particularly students and early-stage researchers, often need several careful readings before the way a paper's core concepts relate to one another becomes clear. This slows down literature review, especially when many papers have to be covered in limited time. This paper presents the design of a Research-Paper-to-Visual-Map Generator, a system that accepts an uploaded research paper and produces an interactive concept map of its key sections, main ideas and the relationships between them. Natural language processing is used to parse the document structure, identify section boundaries and extract key terms and phrases. A relationship-extraction step then links the extracted concepts using contextual proximity and section linkage, and the result is arranged as a hierarchical or network-style map in a web interface, where selecting a node shows the matching excerpt from the original text. The concept extraction and map construction stages were implemented as a prototype and run on the text of this paper to examine the output. The expected outcome is a working tool that shortens initial comprehension and supports faster assessment, note-taking and comparison of research papers.
Keywords: Natural Language Processing, Concept Extraction, Text Summarization, Knowledge Graph, Visual Mapping, Document Structure Analysis, Literature Review Automation, Information Visualization
AI-Based Intelligent Data Deduplication for Cloud Storage
Thanish L.S, Deepak B.T, Dr.Durgadevi.P
Abstract: Cloud storage systems manage amounts of data generated by users, applications and organizations. A significant amount of this data may consist of files or repeated versions of same content, leading to wasted storage space, increased workloads and higher costs for cloud storage. Traditional deduplication methods typically rely on file names, metadata or cryptographic hashes such as SHA-256. While these approaches can identify duplicates, they often fail to capture similar files that contain varying content.This project proposes an AI-Based Intelligent Data Deduplication system for cloud storage, which leverages both hash-based and intelligent similarity analysis techniques. When a file is uploaded, the system generates a fingerprint using cryptography.Additionally, it applies algorithms to identify potentially redundant files.An AI-based decision layer classifies files as unique, exactly duplicated or highly similar.Users can interact with the system through a web interface to upload files, view their storage usage and identify duplicate data.The proposed solution aims to reduce storage utilization and improve storage efficiency while providing users with insights into cloud usage. A working prototype demonstrates that a combination of deduplication methods and sophisticated data analysis techniques are effective for cloud storage management.
SIEM Assistant: Automated Threat Investigation and Reporting Using a Calibrated Random Forest and a Read-Only Advisory Copilot
Amogh K N, Deepak M, Dhanush B R, Harshith H Y, Rajashree M Byalal
Abstract: Security operations centres generate far more alerts than their analysts can adjudicate, and the interval between the first observable event and a defensible verdict is still measured in tens of minutes. This paper presents SIEM Assistant, an investigation and reporting platform that scores unprivileged system telemetry with a calibrated tree ensemble and explains the result through a language interface that holds no execution authority. Ten behavioural attributes — failed authentications, payload volume, Shannon byte entropy, destination port, packet rate, command- and-control reputation, privilege level, command-line encoding, IP reputation and after-hours activity — are extracted in user space and classified by a 100-tree Random Forest over four families. On 600 held-out vectors the classifier reached 98.6% accuracy and a macro F1-score of 0.98 at a mean inference latency of 3.2 ms, while asynchronous ingestion sustained 512 events per second. Calibrating the decision boundary on the deployment base rate reduced false positives by 65% at constant recall and cut mean time to detect from approximately 45 minutes to under three seconds. Response steps are declared as YAML playbooks compiled into directed acyclic graphs and executed in topological order behind an explicit approval record, with technique identifiers mapped to MITRE ATT&CK. The contribution examined most closely here is negative in character: scans that cannot read their inputs terminate in an explicit incomplete state rather than defaulting to a benign verdict, so the classifier is never asked to certify evidence it never saw.
Keywords: Security Information and Event Management, Threat Investigation, Random Forest, Shannon Entropy, Alert Triage, Natural Language Interface, MITRE ATT&CK, Fail-Closed Design.
“An Explainable Al Framework for Multimodal Misinformation Detection and Credibility Assessment”
Ms. Ashwini Baldev Khairnar, Prof. Shivam Limbhare, Prof. Manoj Vasant Nikum
Abstract: The rapid growth of digital media and social networking platforms has significantly increased the spread of misinformation, creating challenges for individuals, organizations, and society. Traditional misinformation detection systems often rely on limited textual features and provide binary predictions without explaining the reasons behind their decisions. This project proposes an Explainable Artificial Intelligence (XAl) framework for multimodal misinformation detection and credibility assessment. The proposed framework is designed to analyze multiple forms of information, including textual content and, where applicable, associated visual or contextual features, to identify patterns that may indicate misleading or unreliable information. Natural Language Processing (NLP), machine learning, transformer-based semantic representations, and multimodal feature extraction are combined to improve detection performance. In addition to classifying content as credible or potentially misleading, the framework generates a credibility score and provides interpretable explanations showing the major factors that influenced the prediction. Explainability techniques are incorporated to reduce the black-box nature of Al-based decision-making and improve user trust. The proposed system can be implemented as an interactive application that accepts user-submitted content and presents the predicted classification, confidence level, credibility assessment, and supporting explanations. The framework can be evaluated using standard performance measures such as accuracy, precision, recall, F1-score, and explainability-related analysis. The overall objective is to develop a transparent, scalable, and research-oriented Al solution that assists users in assessing potentially misleading information while providing understandable evidence for the system's decisions.
Keywords: Explainable Al, Misinformation Detection, Multimodal Al, Natural Language Processing, Credibility Assessment, Transformer Models, Machine Learning, Deep Learning.
Ms. Trupti Sunil Wankhede, Prof. Manoj Vasant Nikum*
Abstract: Modern enterprises manage thousands of servers and infrastructure resources across cloud platforms, making administrative operations increasingly complex and resource intensive. Routine tasks such as virtual machine snapshot creation, backup management, and user access provisioning consume considerable time when executed manually. To address these challenges, we developed InfraAutoX, a cloud-based infrastructure automation framework built using Terraform, Microsoft Azure Cloud Services, Active Directory, Azure RBAC, and devops Pipelines. The framework enables administrators to automate server snapshot creation and user access provisioning through a simple graphical user interface.
The proposed solution follows Infrastructure-as-Code (iac) principles, allowing standardized and repeatable infrastructure operations while minimizing manual intervention. Through backend automation pipelines, administrators can initiate OS and data disk snapshots by entering a server name, while access requests are automatically processed to grant administrator or Remote Desktop Protocol (RDP) privileges through secure integration with Active Directory and Azure Role-Based Access Control.
The framework also incorporates centralized logging, monitoring, and auditing mechanisms to ensure transparency, compliance, and governance. Designed for enterprise-scale environments, InfraAutoX supports large-scale infrastructure management while improving operational efficiency, reducing human errors, enhancing security, and generating significant cost savings. The study demonstrates how cloud-native automation and devops practices can modernize IT operations into a secure, scalable, and cost-effective ecosystem.
Trupti Vinod Marathe, Prof. Manoj V. Nikum, Prof. Mrs. S. N. Raul
Abstract: The rapid proliferation of Internet of Things (IoT) devices in residential and enterprise networks has introduced significant cybersecurity risks, including default credentials, unencrypted communications, and exposed management ports. Traditional perimeter defenses often fail to detect internal subnet vulnerabilities. To address this, we present the IoT Device Security Analyzer, an automated, non-intrusive web application that enhances internal network visibility. Built on a Python Flask framework with an interactive browser dashboard, the system utilizes Scapy-driven ARP scanning and ARP table reconciliation for accurate host discovery. A targeted TCP scanning module performs port auditing and banner grabbing, while a rule-based vulnerability engine identifies high-risk conditions such as cleartext remote shells, exposed TR-069 endpoints. Incorporating MAC OUI vendor fingerprinting and dynamic risk scoring, the analyzer generates prioritized remediation steps and downloadable multi-format audit reports (PDF/CSV/JSON). The tool provides a practical, ethically constrained mechanism for real-time IoT risk mitigation.
Keywords: Internet of Things (IoT), Network Security, Vulnerability Assessment, Device Fingerprinting, Active Scanning, Risk Assessment.
AI Vision-Based Presentation Controller System Using Hand Gesture Recognition
Bhagyashri Dattatray Patil, Prof. Manoj V. Nikum, Prof. Kaustubh C.Bhave
Abstract: Presentations are widely used in educational institutions, business meetings, seminars, and professional environments. Traditional presentation control relies on keyboards, mice, or handheld clickers, which may restrict the presenter's movement. This paper presents an AI Vision-Based Presentation Controller System that enables contactless slide navigation using hand gestures captured through a standard webcam. The system uses Python for application development, OpenCV for video capture and image processing, and MediaPipe Hands for hand-landmark detection. The detected hand landmarks are processed to recognize predefined gestures and map them to presentation commands, including moving to the next or previous slide. Gesture-validation mechanisms help minimize unintended or repeated command execution. The system provides an affordable and accessible alternative to conventional presentation-control devices without requiring specialized hardware. Its performance can be evaluated using gesture-recognition accuracy, command success rate, response time, and false activation rate under different environmental conditions. The system demonstrates the potential of computer vision and hand gesture recognition for natural human-computer interaction.
AI Public Infrastructure Failure Predictor Using Machine Learning
Mr. Kunal Bhatu Kuwar, Prof. Manoj V. Nikum
Abstract: Public infrastructure such as roads, bridges, water pipelines, drainage lines and electrical grid nodes is usually maintained after something has already gone wrong. This paper describes the design and a prototype evaluation of the Public Infrastructure Failure Predictor, a system that combines inspection records, maintenance history, sensor readings, asset age, usage load and weather exposure to estimate how likely each asset is to fail within the next twelve months. Machine learning models (logistic regression, Random Forest and gradient boosting, with XGBoost and LSTM planned for the full system) produce a risk score for every asset, and the main contributing factors are reported alongside it. A Spring Boot backend with JWT-based access control and a PostgreSQL/PostGIS database feeds a React dashboard that shows risk-ranked assets on an interactive map. Because real municipal failure records were not available, the models were tested on a synthetic dataset of 6,000 assets. On held-out data the best model reached a ROC-AUC of 0.83, and ranking assets by predicted risk located 46% of the failures within the top 10% of inspected assets, against about 10% for random inspection. These figures show that the pipeline works end to end, but they must be confirmed on real data before any operational claim is made.
“VoiceGuard: An Explainable AI Framework for Synthetic Voice Authentication”
Ms. Divya Arun Patil, Prof. Shivam Limbhare, Prof. Manoj Vasant Nikum
Abstract: Advances in neural text-to-speech and voice conversion have made synthetic speech increasingly convincing, raising concerns about impersonation and the reliability of voice-based verification. VoiceGuard is proposed as an explainable artificial intelligence framework for assessing the authenticity of recorded speech. The design combines standardized audio preprocessing, Mel-frequency cepstral coefficients (MFCCs), log-Mel spectrogram representations, and a convolutional neural network (CNN) classifier. A separate explanation layer is intended to highlight influential time-frequency regions and provide interpretable evidence alongside a real-versus- synthetic prediction. This paper defines the proposed system architecture, reproducible evaluation protocol, database organization, and analysis criteria. Evaluation on established anti-spoofing benchmarks is planned; no trained-model performance or experimental accuracy is claimed at this stage. The resulting specification provides a practical foundation for subsequent implementation and validation of an auditable synthetic-voice screening application.
NetResilience: Evaluation of Adaptive Recovery Strategies Under Simulated Network Failures
Miss Kajal P. Badgujar, Manoj V. Nikum, Mrs. S. N. Raul
Abstract: Network-dependent applications may experience delays, packet loss, temporary disconnections, and unavailable services. Fixed retry strategies may perform poorly under changing network conditions and can increase response time or create duplicate requests. This paper presents the design of NetResilience, a Python-based simulation system for evaluating recovery strategies under controlled network-failure conditions. The proposed system models network delay, request failure, packet loss, and temporary service unavailability. It compares fixed-delay and exponential-backoff retry strategies with an adaptive retry method that considers observed failures and response-time information. The evaluation framework records request-delivery rate, recovery time, end-to-end latency, duplicate- request count, and retry overhead under different simulated network conditions and workload levels. A web dashboard is proposed to configure experiments and display comparative results. The work is designed as a reproducible software simulation and does not require additional network hardware. Since implementation results have not been supplied for this draft, the paper defines the experimental procedure and evaluation measures without presenting fabricated numerical findings. The completed experiments will determine the conditions under which adaptive retry behaviour offers measurable benefits over baseline strategies.
Mr. Bhushan Sunil Magare, Prof. Shivam Limbhare, Prof. Manoj Vasant Nikum*
Abstract: When critical incidents occur in urban or semi-urban environments, every second counts. Traditional emergency dispatch frameworks continue to rely heavily on manual phone calls and voice descriptions, which frequently lead to location ambiguity, delayed response times, and fragmented coordination. To bridge this gap, we developed SmartResQ, a practical web-based command platform built specifically to streamline emergency reporting and hazard monitoring. The system unites citizen SOS transmissions, a keyword-scoring natural language classifier, deterministic computer-vision heuristics, and simulated Internet of Things (IoT) sensor telemetry into a single cohesive interface. Running on a React and TypeScript frontend backed by a Python Flask 3.0 REST service and an SQLite database, SmartResQ allows users to submit geolocated incident reports while simultaneously monitoring environmental hazards like high temperatures, smoke, and gas leaks. Our implementation successfully tackles common integration pitfalls, such as database alert duplication caused by concurrent frontend-backend triggers. Through local testing and database verification, we demonstrate that combining lightweight heuristics with a structured relational backend offers a robust, responsive foundation for modern smart city emergency management.
EduPulse AI: AI-Based Early Detection of Student Academic Stress Using LMS and Attendance Data
Gurav Aparna Sunil, Mr. Manoj V. Nikum, Prof. K. C. Bhave
Abstract: EduPulse AI is a role-based web application prototype that combines attendance, learning-managementsystem (LMS) engagement, assignment, quiz and academic-performance indicators to support early review of students who may need academic assistance. The system presents three indicator categories—Low, Moderate and High—and provides separate interfaces for administrators, faculty members and students. A synthetic dataset of 2,000 records was used to compare Random Forest, Extra Trees and Gradient Boosting classifiers through stratified five-fold crossvalidation. Extra Trees achieved the highest mean cross-validation accuracy of 0.859 (standard deviation 0.027) and 0.880 accuracy on a held-out set of 400 records. The implementation uses Python, Flask, SQLite, SQLAlchemy and Scikit-learn. Since the target labels were generated using illustrative rules on synthetic data, these results demonstrate the technical workflow only; they do not establish real-world or clinical prediction performance.
Keywords: Learning Analytics, Academic Early Warning, Student Support, Machine Learning, Extra Trees, Attendance, LMS Engagement.
Dr. Beenarani Manoj, Manish Kumar Mandal, Sreejith S, Sanjith Harihar T
Abstract: Choosing the right crop for a field strongly affects yield, resource use and the economic stability of rural households. Conventional expert advice and generalised advisories seldom reach field-level precision, and many digital tools either return a prediction with no explanation or mix model outputs with plain statistical records. This paper presents Smart Crop Recommendation, a web application that keeps the two apart: a machine learning recommendation engine and a regional statistical lookup. The recommendation engine is a Random Forest classifier of 100 decision trees, trained on a 2,200-sample dataset covering 22 crops. It takes seven soil and weather inputs: Nitrogen, Phosphorus, Potassium, pH, temperature, humidity and rainfall. Rather than a single unexplained answer, it ranks every crop and returns the top three with their class probabilities and a comparison of the entered values against recorded values. The lookup component gives a district-level profile for 38 districts of Tamil Nadu, taken from the official statistical handbook (2021–22), and involves no machine learning. On a held-out test set of 440 samples (20% of the data), the model reached an accuracy of 99.32% and a top-3 hit rate of 100.00%. The system also validates user inputs and supports United Nations Sustainable Development Goal 2 (Zero Hunger).
Keywords: Precision Agriculture, Crop Recommendation, Random Forest, Explainable AI, Machine Learning, Agro- Climatic Profile, Sustainable Development Goal 2.
AI-Driven Personalized Career Guidance and Skill Gap Analysis System
P. D. Lanjewar, Gayatri Vijay Patil, Neha Atul Patil, Harsha Ananta Dahivadkar, Rituja Girdhar Bhavsar
DOI: 10.17148/IJARCCE.2026.151102
Abstract: Abstract: The process of selecting a career is challenging, and the general counseling provides limited knowledge about the necessary skills in a particular profession. Our proposed web application identifies the content of a resume in PDF form and uses it to extract its TF-IDF features to classify it using a model to suggest the domain of the career, whether software engineer or data scientist. It then evaluates the skills present in the resume based on the skills required by the job, suggests the skill gaps, and suggests some courses, certificates, and topics that should be studied for interviews. In addition to this, there is also a chatbot, an interview simulator, resume generator, and ATS score checker as other parts of the application. This application is built using Python, Flask, MySQL, and Scikit-learn. All the modules performed as expected in their tests.
Keywords: Artificial Intelligence, Machine Learning, NLP, Career Guidance, Career Forecasting, Skill Gap Analysis, Resume Screening, TF-IDF, Job Interview Preparation, ATS Score
Intelligent Backup Scheduling for Predictive Storage Management
Jatin Verma, Akash Kumar, Abhishek Jatla, Dr. P. Durgadevi
DOI: 10.17148/IJARCCE.2026.151104
Abstract: Data is a central part of everyday digital work. Students, developers, companies, and researchers keep documents, source code, project files, reports, images, databases, and other valuable information on their computers. Such data can be lost through accidental deletion, hard-disk failure, software crashes, malware, power failure, or file corruption, which makes regular backups necessary. Many conventional backup systems follow a fixed schedule, such as a backup every day or every week. This approach is easy to use, but it ignores whether the files have actually changed. A backup may be created even when nothing important has changed, which wastes time, storage space, and system resources. Likewise, if a large amount of data changes right after a scheduled run, the new data stays unprotected until the next one. This project proposes an adaptive backup scheduling system developed in Python. The system monitors a selected source folder and compares its current files with a previous snapshot. New, modified, and deleted files are identified from the file path, file size, and last-modified time. A backup is triggered only when the amount of changed data reaches a configured threshold or when too much time has passed since the last successful backup. CPU usage and free space are also checked before a backup begins. The current prototype uses incremental backup, so only new or modified files are copied to the destination. Each backup is saved in a separate timestamped folder, and a retention policy removes older folders beyond a fixed limit. The project also includes modules for logging, feature generation, training-data generation, and predictive storage analysis, which form a base for future machine-learning integration. The proposed system aims to make backups more efficient, reliable, and responsive to real changes in data.
TWO-WAY INDIAN SIGN LANGUAGE COMMUNICATION ASSISTANT USING MACHINE LEARNING
Rashmi R, Bhavana L, Sarah Ninan, Adithya Pratheep, Dr. Beenarani Manoj
DOI: 10.17148/IJARCCE.2026.151105
Abstract: It can be sometimes a tough challenge for the Indian Sign Language (ISL) user to communicate with the non- ISL learner. In this regard, we decided to address the issue by creating a Two Way Indian Sign Language Communication Assistant using Machine Learning. Instead of simply converting the sign to speech, our system focuses on both sides of communication. The system also provides the extra functionality for ISL learners. A webcam is used for capturing the sign gestures. The points of interest for hands, face and body are identified using MediaPipe Holistic. All these points are taken under consideration for recognizing the gestures through LSTM model analysis of multiple video frames. Signs recognized are then converted to text and spoken. In case of communication by speech, it takes input from the microphone, converts to text, and shows corresponding signs in the form of images/videos/animations. Apart from this, there is a learning mode, where one can practice and get familiar with the sign gestures of ISL. All these features combine sign recognition, speech conversion and learning of sign language into a single application. The goal of this project is to ease the communication between ISL users and non-signers and also to create more awareness about the language.
Keywords: Indian Sign Language , Sign Language Recognition, Two-Way Communication, Speech-to-Sign Translation, LSTM, MediaPipe Holistic, Gesture Recognition, Human-Computer Interaction, Assistive Technology, Sign Language Learning
Abhijith G, Jijohns Johnson, Vishnu S, Vishakh V Prakash, Anpu Alexander, Girija V R
DOI: 10.17148/IJARCCE.2026.151106
Abstract: Malware classification is an important cybersecurity task because identifying malware families can help analysts understand threats and support appropriate response. This literature survey reviews sixteen studies addressing malware classification and detection through image-based learning, one-dimensional signal representations, feature selection, memory-dump analysis, multimodal fusion, hybrid deep learning, and Android-focused methods. Particular attention is given to binary-to-image conversion and convolutional neural networks (CNNs), which learn visual patterns from executable bytes without requiring conventional disassembly. The reviewed studies report promising results, including accuracy above 90% in several image-based settings and 99.05% accuracy for an epoch-selection ensemble on the Malimg dataset. However, reported results are not directly comparable because studies use different datasets, class definitions, representations, and evaluation protocols. Recurring challenges include dataset bias, generalization to unseen malware, class imbalance, information loss during representation and resizing, model complexity, and the trade-off between static analysis and execution-dependent evidence. The survey identifies the need for careful evaluation of binary-to-image preprocessing, lightweight CNN architectures, reproducible comparisons, and validation across datasets. These findings provide a basis for the project’s static, image-based malware-family classification direction while leaving final model configuration to empirical evaluation.
TERRA: Topological Extraction for Rapid Response Analysis – A Literature Survey
Adithya Krishna, Athifa Roshni, Harsha N S, Sooraj S Nair, Sumi Rehim, Girija V R
DOI: 10.17148/IJARCCE.2026.151107
Abstract: Up-to-date road networks are essential for rescue and relief operations after earthquakes, floods and wildfires, yet roads are often unmapped, occluded or damaged when they are needed most. This paper presents a literature survey carried out as the groundwork for TERRA (Topological Extraction for Rapid Response Analysis), a project on extracting road networks from satellite imagery for rapid response. Sixteen recent studies are reviewed and organised into five themes: encoder–decoder segmentation networks, graph- and topology-based road extraction, rural and emergency road mapping, disaster damage assessment, change detection and road condition monitoring, and disaster-aware path planning. For each study the methodology, reported results and limitations are summarised and compared. The survey shows that segmentation networks achieve good pixel-level accuracy but fragment narrow and occluded roads, that graph-based methods improve topological correctness at higher cost or with extra prerequisites, and that few works connect road extraction with damage-aware routing. Common research gaps are consolidated and a direction for TERRA is outlined.
Abstract: Tracking apps that run on a phone stop working once a thief switches the handset off or cuts its connection, so the owner loses both the device and any way of reaching it. This paper presents an anti-theft ecosystem in which recovery is handled by a separate companion tracker with a battery of its own, a dedicated SIM card and a built-in satellite receiver. The tracker combines an ESP32-class controller, a GPS receiver and a SIM800L GSM/GPRS modem. It sends its position to an online dashboard, replies to location requests received as text messages, issues an SOS alert carrying a map link when a button is held, and accepts calls from the owner. A temporary Wi-Fi page stores the device identity, API key, access point name, upload interval and emergency numbers in flash memory, so no secret is written into the program. A Bluetooth Low Energy layer that detects separation from the phone is specified as the planned extension. The prototype was exercised through its configuration portal and its text-message interface; the observed behaviour, including a location reply carrying six satellites and an SOS alert delivered with a map link, is reported. Weaknesses of the present firmware and the next development steps are also discussed.
Keywords: Bluetooth Low Energy, GSM, GPS, anti-theft, offline tracking, SOS alert, ESP32, SIM800L, smartphone recovery.
AI-Driven Dynamic Wireless EV Charging System: Battery Health Prediction and Energy Optimization Using IoT Data
Ashna Akbar Pinjari, Nilesh Bhila Patil, Nilesh Rajendra Pardeshi, Janhavi Rajendra Mahajan, P. V. Baviskar
DOI: 10.17148/IJARCCE.2026.151109
Abstract: Charging is still the weakest part of owning an electric vehicle (EV). Sessions are long, public chargers are unevenly spread, and the battery tends to wear out sooner than owners expect. Most published work tackles only one of these problems: the wireless power link, the estimation of battery health, or the sensing of battery data. This paper describes the design of a dynamic wireless charging system that brings these pieces together and adds an energy- optimization step. Transmitter coils placed under the road supply power to a receiver coil on a moving vehicle by inductive coupling. An ESP32-based sensing unit records voltage, temperature, charge cycles and charging time and sends them to a cloud platform, where machine learning models are intended to estimate battery health and suggest changes to the power transfer. A web dashboard presents the results to the operator. The work reported here covers the first project stage: system architecture, component selection, preparation of a battery dataset and exploratory analysis. On a sample of ten records, battery health dropped from 95% to 87% between cycles 5 and 14 (r = -0.98), and higher temperature went with lower health (r = -0.63). The planned prediction models and the remaining hardware validation are also outlined.
Keywords: Dynamic wireless power transfer, electric vehicle, battery health prediction, Internet of Things, machine learning, energy optimization
Abstract: Meetings are an important part of academic and professional work, but the information discussed during them is not always recorded in a consistent or useful form. Preparing minutes manually can take time, and important decisions, responsibilities, deadlines, and risks may be overlooked. This paper presents MeetMind, an AI-powered meeting summarization and action-item management system. The application accepts either a typed meeting transcript or an uploaded audio/video recording. For recordings, it generates a transcript using the OpenAI speech-to-text service when configured, or a local Whisper-based transcription option. The transcript is then processed by an offline natural language processing pipeline that uses spaCy, NLTK, and scikit-learn. TF-IDF sentence ranking supports extractive summaries, while linguistic patterns and named-entity information assist in identifying decisions, tasks, owners, deadlines, priorities, and risks. An optional locally hosted language model through Ollama can provide additional summary and insight generation. MeetMind stores meeting records in an SQLite database with FTS5 full-text search. The stored information supports a Memory Assistant, an Accountability Center, calendar-event export, analytics, a Meeting Readiness and Dependency Radar, and report generation. The prototype was tested at module level using eight functional test cases, all of which passed. In one sample product-strategy meeting containing 617 words and eight speakers, the system identified 15 action items and eight decisions and produced a productivity score of 86.5 out of 100, displayed as 87%. These observations demonstrate the operation of the prototype; they do not represent validated accuracy measurements. The paper describes the system design, implementation, observed outputs, limitations, and directions for further evaluation.
Abstract: Wireless sensor networks require intrusion detection methods that respect limited memory, processing capacity, bandwidth, and battery supply. This paper presents an implemented multiclass Denial-of-Service (DoS) attack detection system for Wireless Sensor Networks using the WSN-DS dataset and lightweight machine-learning models. The public prototype exposes manual prediction and bulk CSV analysis through a React front end and a Flask API. The application reports 374,661 dataset records, 17 input features, five attack classes, and a headline accuracy of 99.47%. Its model-comparison dashboard reports the highest displayed accuracy for a 17-feature Random Forest at 99.72%; however, the Random Forest is not the most time- or memory-efficient option. The study evaluates the full 17-feature representation and also selects a reduced Decision Tree using the five highest ranked features: Is CH, ADV S, SCH S, Data Sent To BS, and Expanded Energy. The five classes are Normal, Flooding, Blackhole, Grayhole, and TDMA. The selected five-feature Decision Tree provides the intended balance of accuracy, training efficiency, model size, interpretability, and reduced feature processing overhead for a lightweight WSN detector. Training-time and memory values are reported as displayed by the dashboard and are not interpreted as packet level inference latency or direct energy measurements. The study links a reproducible web deployment to a resource-aware WSN DoS detection design and identifies the measurements required for hardware-level validation.
Abstract: Artificial intelligence (AI) and biometric recognition technologies have created new possibilities for automated identity verification. However, applying these technologies in an election environment requires careful consideration of recognition errors, presentation attacks, data privacy, processing latency, and human accountability. This paper presents a proposed framework for real-time duplicate-voter screening during Election Day polling using facial recognition, a same-day voter registry, voice-assisted confirmation, and human-in-the-loop review.The proposed framework captures a facial image, generates or extracts facial features using the selected recognition method, and compares the observation against eligible records in the current-day registry.
Its proposed contribution is an end-to-end polling-day screening workflow that connects temporal registry search, multimodal evidence, uncertainty handling, and human review. No performance improvement is assumed without experimental evidence. The framework is intended as a research prototype and does not replace official electoral verification or authorized human decisions.[3]
A Survey on Recent Research Trend Towards Adaptive Reliable Intelligent Scalable Processing Engine for Sustainable Fault-Tolerant Embedded Systems
G. Soujanya, Ranganath. B, Adithya. D. A, Geetha. V, Sachin. M. H, T. S. Koushik
DOI: 10.17148/IJARCCE.2026.151113
Abstract: Conventional fault-tolerant techniques rely on fixed redundancy or protect isolated blocks, increasing area, power, and complexity while offering limited fault classification and adaptive recovery. This paper presents ARISE, a dedicated supervisory processor for sustainable fault-tolerant embedded systems. It combines a 32-bit multi-cycle core with fault detection, classification, severity assessment, reliability monitoring, result verification, and adaptive recovery with safe escalation. Implemented in synthesizable Verilog, the design is verified through RTL simulation and fault injection using Cadence tools, synthesized with Cadence Genus, and demonstrated in an automotive wheel-speed monitoring application. ARISE provides a scalable foundation for reliability-critical systems.
Mubeena Banu, Chandana K, Lavanya N L, Nayanashree D
DOI: 10.17148/IJARCCE.2026.151114
Abstract: The Phishing Website Identification System is designed to detect and prevent fraudulent websites that attempt to steal sensitive user information. Phishing attacks have become a major cybersecurity threat in the digital world. These attacks trick users into visiting fake websites that resemble legitimate platforms. The proposed system uses Machine Learning techniques to identify phishing websites. It analyzes various URL features, domain information, and website characteristics. The extracted features are processed and provided to a trained machine learning model. Algorithms such as Random Forest, Decision Tree, and Support Vector Machine can be used for classification. The system classifies the submitted website as either legitimate or phishing. It provides quick and automated detection of suspicious websites. The proposed approach aims to improve detection accuracy and reduce false-positive rates. It also helps identify potentially harmful websites and alerts users about security risks. This system can be integrated into web applications or browser- based security tools. Overall, the project aims to provide a reliable and intelligent solution for improving online safety and protecting users from phishing attacks.
Abstract: Modern Artificial Intelligence (AI) applications use specialized models to perform different tasks. However, selecting a single model for all user queries may result in reduced performance when queries vary in intent and complexity. To overcome this challenge, this project proposes an “AutoML Framework for Intelligent Model Selection,” a dynamic multi-agent system consisting of five specialized AI agents: Text Model,Embedding Model, Vision Model, Reasoning Model, and Code/Data Model.
These agents collaboratively analyze incoming user queries and identify the most suitable model based on the requirements of each query.The framework incorporates an Automated Machine Learning (AutoML) evaluation engine using five algorithms: K-Nearest Neighbors (KNN), Recurrent Neural Networks (RNN), Linear Regression, Logistic Regression, and Naive Bayes. The system evaluates agent suitability using relevant query features and performance metrics, including Receiver Operating Characteristic (ROC) curves and Area Under the Curve (AUC) scores based on labelled evaluation data. The most suitable agent is selected to generate the final response. An interactive frontend dashboard displays agent scores, algorithmic metrics, selection results, and the reason behind the selected model. The proposed framework aims to provide intelligent, adaptive, and transparent model selection for different types of AI queries.
Keywords: AutoML, Intelligent Model Selection, Multi-Agent Systems, Text Model, Embedding Model, Vision Model, Reasoning Model, Code/Data Model, KNN, RNN, Linear Regression, Logistic Regression, Naive Bayes, ROC-AUC, Visualization
SignAI: Intelligent Bidirectional Indian Sign Language Translation Using Skeleton Keypoints
Anand S. Balapure, Vedant Y. Badgujar, Netra A. Jayaswal, Tejaswini B. Sharma, Sheetal P. Patil
DOI: 10.17148/IJARCCE.2026.151116
Abstract: Communication between Deaf and hearing individuals remains difficult because interpreters and accessible translation tools are scarce. This paper presents SignAI, a bidirectional Indian Sign Language (ISL) translation prototype that combines Sign-to-Text (S2T) and Text-to-Sign (T2S) translation in a single web application. The S2T module classifies isolated signs from MediaPipe Holistic keypoints using a Conv1D–BiGRU–attention network, supported by motion-based segmentation and confidence gating; recognised glosses are turned into English sentences with a rule-based generator or few-shot prompting of FLAN-T5-base. The T2S module maps text to validated glosses and renders the matching sign clips as a skeleton video, with optional PoseTransformer and fingerspelling support. On a custom 60-class ISL adjective dataset of 1,022 clips, the classifier reaches 74.29% top-1 and 88.57% top-5 accuracy on the test set (n = 140; 95% CI: 66.47–80.81%). At a confidence threshold of 0.65, 89.16% of accepted predictions are correct at 59.29% coverage. A data and code audit and its corrective measures are also reported.
Keywords: Sign language recognition, Indian Sign Language, MediaPipe, skeleton keypoints, bidirectional GRU, sign language production, confidence gating.
Abstract: Surveillance cameras are only useful if they keep working when conditions turn bad: at night, in wind, under shifting shadows. We built a compact surveillance prototype on a Raspberry Pi 5 that pairs an infrared-capable camera with classical motion analysis and a small deep learning detector. Each frame goes to YOLOv8n first, which looks for people. Frames with no person are passed to a motion analysis stage, where OpenCV background subtraction and noise filtering pick out moving regions, and anything that moves but isn't recognised is flagged as a possible anomaly. An operator can label flagged objects so the model can be retrained later. We also compared YOLOv8n against MoveNet SinglePose for human detection and found YOLOv8n faster and steadier on this hardware, and running the detector through OpenVINO raised the frame rate on the Pi from about 1 to about 6 FPS. This paper describes the design, the reasoning behind each choice, and the weaknesses we ran into.
Smart Flood Detection and Impact Mapping System Using GSM-Based Alerts and Satellite Image Heat Maps
Mrs. Vyshnavi V J, Amar, Divit V, Hari Narayana S, Kiran S
DOI: 10.17148/IJARCCE.2026.151118
Abstract: Floods rank among the most damaging natural hazards, and the harm to people and property grows sharply when warnings come late or never reach the affected community. Most IoT flood monitors described in recent literature depend on internet links, cloud back-ends and smartphone applications, which are exactly the services that become congested or fail during a disaster; moreover, they seldom assist with damage assessment after the water recedes. This paper presents a low-cost, two-stage framework that addresses both phases. The first stage is an Arduino-based field node that reads a water level sensor and a rain sensor, grades the situation as Normal, Warning or Danger against fixed thresholds, drives local LED and buzzer indicators, and, at the Danger grade, sends SMS messages through a SIM800L GSM modem without needing the internet. A battery or solar supply allows deployment at remote sites. The second stage is a Python pipeline that compares pre- and post-flood satellite images, isolates newly inundated regions, assigns severity classes and renders red-orange-green heat maps together with the affected area in km². A bench prototype using a water container demonstrates the sensing and indication logic. The framework supports early warning for residents and farmers and provides authorities with a spatial view for planning relief.
Keywords: Flood detection, Internet of Things, Arduino, GSM SMS alert, satellite imagery, heat map, disaster management.
Real-Time Coding Contests in College Programming Labs: A Case Study at RCPIT Shirpur
Sachin Narendra Chaudhari, Jayesh Rameshsing Girase, Sakshi Dadabhau Patil, Chetan Prabhakar Sonawane, Dr. N. G. Shinde
DOI: 10.17148/IJARCCE.2026.151119
Abstract: Data Structures and Algorithms (DSA) is a core subject in undergraduate computer engineering and artificial intelligence curricula. In most engineering colleges, programming practicals follow an open-ended lab format where students work at their own pace and submit unfinished tasks as homework. Under this model, instructors frequently observe three challenges: students postpone coding until the weekend deadline, make repeated trial-and-error submissions instead of thinking through logic, and struggle on proctored practical exams. To address these problems, we created and deployed Campus Lab, an open-source web platform for college computer laboratories. Campus Lab brings real-time, timed contest rooms and sandboxed code execution (using Judge0 CE) directly to the departmental local area network (LAN). Instructors run 90-minute contests during standard two-hour practical sessions, offering instant feedback on sample tests and a live room leaderboard. We evaluated Campus Lab in a 16-week crossover study with 64 sophomore Artificial Intelligence and Machine Learning students at R. C. Patel Institute of Technology (RCPIT), Shirpur, split into two batches (n = 32 each). Students in contest mode completed 88.5% of assigned problems during lab hours, compared to 61.2% in traditional homework mode. On proctored practical exams, students scored 5.8 marks higher (out of 50) on topics learned through contest rooms. Formal submissions per solved problem dropped from 5.7 to 3.2, as students made use of sample test runs and planned their logic before submitting. Survey responses showed strong student enthusiasm for contest-driven labs when grading prioritized problem completion rather than leaderboard speed. We share four practical recommendations for engineering colleges adopting in-class contest platforms.
StudyMate: A Multi-Agent Academic Assistance System
S. L. Tambe, Harshada Vinod Jagdale, Mrunal Madhukar Bharat, Jidnyasa Deepak Patil, Srushti Mahesh Patil
DOI: 10.17148/IJARCCE.2026.151120
Abstract: The rapid expansion of academic content and digital learning resources requires centralized, secure, and intelligent platforms to assist engineering students in their daily studies and placement preparations. Traditional academic portals often suffer from fragmented file management, rigid access controls, lack of personalized branch-wise filtering, and absence of real-time AI assistance. This paper presents StudyMate: A Multi-Agent Academic Assistance System, a modern, web-based platform tailored exclusively for students and administrators of the R. C. Patel Institute of Technology (RCPIT). StudyMate consolidates all essential academic resources into a centralized, lightning-fast Single Page Application (SPA). The frontend is engineered using React.js, Vite, and Tailwind CSS, offering a modern, responsive user interface. The backend is powered by a robust Python Flask RESTful API interacting with a cloud-hosted MongoDB Atlas database. Security is enforced through JSON Web Tokens (JWT) for stateless session management and Werkzeug for cryptographic password hashing. Furthermore, StudyMate integrates advanced functional modules including a dynamic branch-and-semester resource filter, an interactive digital library, placement aptitude mock tests, a curated Final Year project repository, and a 24/7 AI-driven doubt-solving Hub. To ensure smooth institutional management, the platform features a strict Exclusive Admin Architecture incorporating credential tracking and real-time student account impersonation. Comprehensive testing confirms the stable, secure, and rapid execution of all client-server workflows.
Keywords: Academic Resource Platform, React.js, Python Flask, MongoDB Atlas, JSON Web Tokens (JWT), Role- Based Access Control, AI Doubt Solver.
AI Based Real-Time Data Analysis and Decision Generation System for Gokul Text Print
Dr. Pankaj V. Baviskar, Hemanshu Umesh Suryawanshi, Priyanshu Dipak Patil, Lakshvedhi Anil Patil, Gaurav Manohar Patil
DOI: 10.17148/IJARCCE.2026.151121
Abstract: Manufacturers rely on enterprise resource planning (ERP) systems to track sales, inventory, production, and order fulfilment, yet the records often sit in separate modules and need heavy manual reading before they can guide decisions. This paper presents an AI decision-support platform for Gokul Text Print that links ERPNext data with machine-learning analytics, explainability, retrieval-augmented generation (RAG), and specialized AI agents. A Frappe REST API client and a validation-focused extraction, transformation, and loading (ETL) pipeline pull and check the records. The sales module classifies decision direction and derives probability-based projections, the inventory module estimates reorder risk, SHapley Additive exPlanations (SHAP) support model interpretation, and a Sentence Transformers and Facebook AI Similarity Search (FAISS) retriever supplies organizational knowledge for grounded answers. A manager agent combines the outputs of sales, inventory, and knowledge agents before a language model writes the final response, and a Flask dashboard shows indicators, recommendations, alerts, and reports. Preliminary development runs gave 68.45% accuracy and a receiver operating characteristic area under the curve (ROC-AUC) of 0.7447 for sales, and 100% accuracy with a ROC-AUC of 1.0000 for inventory, a result that still needs temporal and leakage checks. The contribution is an auditable integrated architecture rather than a new algorithm, together with an evaluation framework for prediction, retrieval, agent quality, and reliability.
A Comparative Study of Machine Learning Algorithms for Network Intrusion Detection
Akash jain, Dr. Pradeep Yadav
DOI: 10.17148/IJARCCE.2026.151122
Abstract: The increasing sophistication of cyberattacks poses critical challenges to modern network infrastructure security. Traditional signature-based intrusion detection systems (IDS) are becoming obsolete against zero-day exploits, polymorphic malware, and advanced persistent threats (APTs). This paper presents a comparative study of machine learning (ML) algorithms for network intrusion detection, evaluating Random Forest (RF), Gradient Boosting (GB), and Logistic Regression (LR) on both binary (normal vs. attack) and multiclass classification of nine traffic categories: Normal, Denial-of-Service (DoS), Probe, Remote-to-Local (R2L), User-to-Root (U2R), Botnet, Brute Force, SQL Injection, and Cross-Site Scripting (XSS). Experiments use a synthetic, NSL-KDD-inspired dataset of 4,506 labeled network traffic samples with 35 engineered features, evaluated with a stratified 80/20 hold-out split and five-fold stratified cross-validation. All three algorithms achieve 100% accuracy in binary classification, while Random Forest achieves 90.91% accuracy and a weighted F1-score of 0.9114 in multiclass classification. Per-class analysis shows that the web-application attacks (Brute Force, XSS, and SQL Injection) are the hardest to separate at the network-flow level. Feature importance analysis shows that behavioral indicators such as num_compromised, src_bytes, num_file_creations, and num_failed_logins are the strongest predictors of malicious activity. The results are compared with figures reported in the literature, and the study discusses the practical trade-offs among accuracy, interpretability, and computational cost when selecting an algorithm for deployment.
CounselAI: A Multi-Agent Conversational Counsellor with Calibrated Cutoff Forecasting for MHT-CET Engineering Admissions
Dr. Nitin Shinde, Dnyaneshwari Ganpatsingh Girase, Dhiren Sanjay Thakur, Om Ujjwal Patil, Vivek Narendra Patil
DOI: 10.17148/IJARCCE.2026.151123
Abstract: Every year, students who clear the Maharashtra Common Entrance Test (MHT-CET) have to choose and order about strategy. In this paper we present CounselAI, a chat-based counselling system for these students. A large language model (LLM), coordinated through LangGraph, works with six agents that handle the student profile, eligibility, ranking, CAP-round strategy, document checklists and the final explanation. The LLM never calculates any number itself: cutoffs and chances come from a statistical model, and a grounding check rejects any reply that contains a number the tools did not produce. We built a dataset of 92,135 official closing-percentile records from nine Directorate of Technical Education (DTE) cutoff documents for 2022 and 2023, covering 338 colleges, 1,921 college–branch programs, 81 seat-category codes and three CAP rounds. On colleges that were kept out of training, our gradient-boosted model predicts the next year’s cutoff with a mean absolute error (MAE) of 8.10 percentile points, 18.0% lower than simply reusing last year’s cutoff (9.88), and the error drops to 1.53 points for programs with cutoffs in the 90–100 band. The admission chances derived from it are well calibrated (Brier score 0.058, AUC 0.976, decision accuracy 91.97%), and students labelled “Safe” were admitted 96.4% of the time. For deciding whether to accept a seat or wait for the next round, a classifier trained on 2022 and tested on 2023 reached 70.9% accuracy and 64.5% precision, compared with 66.4% and 54.7% for a simple rule based on the previous year. The system replies in English or Marathi. On 16 held-out test messages in English, Hindi-English and Marathi, Gemini read the profile correctly for 97.5% of fields, all 31 explanations it wrote passed the grounding check, and the median reply time was about 4 s. A rule-based fallback (98.7%) keeps the system working when the LLM is unavailable. CounselAI is deployed as a public web application. Our results suggest that keeping an LLM counsellor tied to a calibrated statistical model makes its advice something that can be checked rather than taken on trust.
Keywords: admission counselling, multi-agent systems, large language models, tool-grounded generation, cutoff prediction, gradient boosting, probability calibration, MHT-CET, Marathi language.
PlaceNext: A Centralized Hub for Institutional Talent Acquisition
Mihir Tathe, Lokesh Patil, Yash Marathe, Kuldip Mahale, Prof. S. L. Tambe
DOI: 10.17148/IJARCCE.2026.151124
Abstract: Campus placement is one of the most important activities of a technical institution, yet in many colleges it is still managed through spreadsheets, messaging groups, paper forms, and e-mail. Student data is scattered across files, eligibility is checked by hand, drive announcements reach students unevenly, and the placement cell has no single, reliable record of who applied, who was shortlisted, and who was placed. This paper presents PlaceNext, a web-based placement management application that brings these activities into one centralized system. The application provides four role-based logins: Super Admin, Training and Placement Officer (TPO), Student, and Recruiter. The Super Admin manages TPO accounts and system settings; students maintain a verified profile and resume; recruiters register and post job openings directly in the system; the TPO reviews and approves each job and publishes it as a company drive with eligibility criteria; an eligibility engine matches students to drives using rules such as minimum CGPA, permitted backlogs, branch, and graduation year; and applications are tracked through each selection round until the final result. Notifications, interview scheduling, placement analytics, and downloadable reports support follow-up and decision- making. All records are stored in one central database so that the placement cell, students, and recruiters work from the same information. The prototype was evaluated at module level using 12 functional test cases, of which all 12 passed, and was demonstrated with 120 students and 6 company drives. These observations demonstrate the operation of the prototype; they do not represent validated accuracy or user-satisfaction measurements. The paper describes the system design, implementation, observed outputs, limitations, and directions for further evaluation.
Keywords: Placement management system, campus recruitment, training and placement cell, eligibility engine, role- based access control, web application, placement analytics.
AI Agentic Supply Chain Intelligence System: Real-Time GPS Tracking, ETA and Delay Prediction, and AI-Driven Recommendations for Logistics Monitoring
Dr. N. A. Patil, Yash Pradip Patil, Himanshu Sudhir Mahajan, Gaurav Kiran Patil
DOI: 10.17148/IJARCCE.2026.151125
Abstract: Online shopping, global trade and expanding transportation networks have made logistics monitoring a central concern for modern businesses. Most vehicle-tracking systems only show where a vehicle is at a given moment; they do not predict delays, estimate arrival times reliably, or support decisions. This leads to higher costs, late deliveries and unhappy customers. This paper presents the AI Agentic Supply Chain Intelligence System, a web-based platform that combines real-time GPS tracking, route visualization, estimated time of arrival (ETA) prediction, machine-learning- based delay prediction, fleet monitoring and AI-generated recommendations in one dashboard. The system analyses vehicle location, route distance, traffic level and weather risk, and uses Random Forest regression to forecast shipment delays. The platform is built with Python, FastAPI, Scikit-Learn, OpenStreetMap, OpenRouteService, HTML, CSS, JavaScript and Leaflet.js. A working prototype was developed and validated through eight functional test cases covering tracking, routing, ETA, delay prediction, fleet monitoring, recommendations, analytics and reporting, all of which passed. The results show that combining geospatial technologies with predictive analytics improves supply chain visibility and supports faster, better-informed logistics decisions.
Keywords: Supply Chain Intelligence, GPS Tracking, ETA Prediction, Delay Prediction, Random Forest, Route Visualization, Fleet Monitoring, FastAPI, OpenStreetMap, AI Recommendation
FITGUARD AI: A RULE-BASED FITNESS AND NUTRITION DECISION SUPPORT SYSTEM
Khatik Mohammad Sabique Sikandar, Mohit Manilal Patel, Bhavesh Dinesh Patil, Pranav Kishor Sutar, Prof. Kaminee Subhash Patil
DOI: 10.17148/IJARCCE.2026.151126
Abstract: Generic diet and exercise advice ignores differences in body size, daily activity, goals and food preferences, while hand-made plans are slow to prepare. This paper describes FitGuard AI, a web-based decision support prototype that converts a short user profile into a personalized calorie target, a seven-day meal plan, a weekly workout split and a screening-level risk indicator. The back end is written in Python with Flask and keeps profiles and audit entries in SQLite; the front end uses HTML, CSS and JavaScript. Energy needs are estimated with the Mifflin–St Jeor equation, scaled by an activity factor and shifted by a goal-specific adjustment, and every result is accompanied by an explanation trace that lists the formulas and rules that produced it. In a recorded example (male, 20 years, 180 cm, 54 kg, active, muscle-gain goal) the system produced a BMI of 16.7, a BMR of 1570 kcal/day, an energy expenditure of 2708 kcal/day and a target of about 2960 kcal/day. All five modules passed the functional tests of the first project stage. The recommendation path is deterministic; no trained learning model or image analysis is implemented, and no clinical validation has been done. The paper therefore presents the system as an inspectable engineering baseline, discusses its safety and privacy limits, and lists the evaluation needed before real-world use.
Varad Sane, Varad Kulkarni, Shrikant Somani, Prashant Yanpallewar
DOI: 10.17148/IJARCCE.2026.151129
Abstract: Agricultural production is influenced by spatial and temporal variations in crop health, soil moisture, temperature, humidity, and weather conditions. Conventional field inspection and fixed-schedule irrigation may not adequately represent these variations when decisions are required for individual crop zones. This paper presents the design and software implementation of an integrated precision crop-farming decision-support framework combining plant-image classification, weather information, configurable decision rules, and a software-based digital twin. The system employs MobileNetV3-Large as its plant-disease classification backbone. A React and Vite frontend communicates with a Java Spring Boot REST API, which coordinates a Python-based PyTorch inference service, PostgreSQL persistence, and the Open-Meteo weather service. The decision engine combines available crop-zone state, classification output, confidence score, and forecast context to generate one of four recommendations: Irrigate, Treat, Monitor, or Skip. A simulation component records virtual state transitions and maintains a history of decisions and actions. The proposed model-training configuration uses ImageNet-pretrained weights, two-phase fine-tuning, and PlantVillage as the baseline dataset, with PlantDoc identified for external robustness evaluation. The current work establishes a software workflow for integrating crop-health evidence and environmental context. Quantitative model evaluation, field validation, physical actuation, and measured resource-saving experiments have not yet been completed. The framework therefore provides an implementation baseline and a reproducible evaluation pathway for future precision-agriculture research.
Keywords: Precision agriculture, plant disease classification, MobileNetV3-Large, weather-aware irrigation, decision- support system, agricultural digital twin.
NEOARCHITECT: A LOCAL AI-ASSISTED ARCHITECTURAL DESIGN AND SYNTHESIS PLATFORM WITH PRIVACY-PRESERVING STRUCTURAL INTELLIGENCE
Afsal S, Aravind A, Salini Lalu, Abhishek Babu, Merlin K Thomas, Jinu L
DOI: 10.17148/IJARCCE.2026.151130
Abstract: Architectural design requires coordinated decisions about room relationships, circulation, geometry, structural safety, regional aesthetics, materials, cost, and visualization. Conventional computer-aided design workflows often involve repeated manual drafting, while image-oriented generative systems may produce visually plausible plans without preserving editable geometry or engineering context. This paper presents NeoArchitect, a proposed standalone and offline AI-assisted platform that connects natural-language requirement parsing, architectural-style classification, adjacency- constrained two-dimensional vector floor-plan generation, preliminary structural and seismic screening, material and cost estimation, and parametric three-dimensional synthesis. The system is designed to keep project data and model inference on the local machine, reducing dependence on cloud services for core processing. The Phase-I work establishes the problem context, reviews 16 core research works as represented in the project survey, identifies privacy, engineering integration, and regional-style representation as research gaps, and defines a modular architecture and validation strategy. The paper describes the proposed workflow and its key modules. Since the project is at the Phase-I design and planning stage, no measured model accuracy, end-to-end runtime, or validated structural performance is claimed. Engineering outputs are intended only for preliminary decision support and require independent professional verification.
Keywords: Local AI, architectural design, floor-plan generation, natural-language processing, graph-based layout, structural intelligence, privacy-preserving AI, 3D visualization.
Abstract: The exponential growth of digital content in multiple formats — text documents, audio recordings, and video lectures — has created significant challenges for efficient information extraction and knowledge synthesis. This paper presents InsightAI, a comprehensive web-based application that leverages Large Language Models (LLMs) and Natural Language Processing (NLP) techniques to provide intelligent multi-modal document summarization, interactive question answering, and analytical insights. The proposed system accepts inputs in seven formats including PDF, DOCX, TXT, MP3, WAV, MP4, and YouTube URLs, and utilizes the Google Gemini 2.5 Flash API for text processing and the Groq Whisper Large-V3 API for audio transcription. InsightAI offers six distinct summarization modes, a Retrieval- Augmented Generation (RAG) based QA chatbot, sentiment analysis, keyword and Named Entity Recognition (NER) extraction, automated quiz generation, flashcard creation, mind map visualization, and multi-language translation across ten languages. Summary quality is scored by a reference-free, LLM-based judge for coherence, consistency, fluency and relevance. A multi-user layer provides email-verified registration with one-time passwords (OTPs) sent through Nodemailer, bcrypt-protected passwords, signed session tokens, password recovery, and a relational database (SQLite for development, PostgreSQL for production) that stores each user’s summaries and chat history privately. The system is built using a Python FastAPI backend and a React.js frontend, providing a scalable and user-friendly platform for students, researchers, and professionals.
Keywords: Large Language Models, Natural Language Processing, Document Summarization, Question Answering, Retrieval-Augmented Generation, Multi-Modal Processing, LLM-as-a-Judge Evaluation, User Authentication, One- Time Password, Sentiment Analysis, Google Gemini, Groq Whisper
Akshat Kumar, Rudrakash Agarwal, Varun K, Dr. Beenarani Manoj
DOI: 10.17148/IJARCCE.2026.151137
Abstract: AeroPredict-X is a flight-delay prediction and analytics system built using United States flight records covering January 2024 to December 2025. The project standardizes monthly datasets and applies supervised learning to two tasks: classifying whether arrival delay reaches 15 minutes and estimating arrival-delay duration. A CatBoost-based classifier is evaluated with chronological data partitions and metrics suited to imbalanced classification, while regression performance is measured using error-based metrics. The reported test results include ROC-AUC of 0.6671 and PR-AUC of 0.3601 for classification; the displayed threshold of 0.24 gives an F1-score of 0.3805. The regression model reports a mean absolute error of 26.60 minutes and R² of 0.0164, indicating limited explanatory power. SHAP explanations and a Streamlit interface support interpretation, historical analytics, single predictions, and validated batch CSV scoring. AeroPredict-X is a prototype, not an operational guarantee; further validation, calibration, leakage checks, and real-time data integration are needed.
EMPLOYMENT STATUS CLASSIFICATION USING SUPPORT VECTOR MACHINES OF VISUALLY-IMPAIRED GRADUATES
Neelesh N Shrinivisan, Aparna M, Hasanthika Sri S, G. Paavai Anand
DOI: 10.17148/IJARCCE.2026.151148
Abstract: This study uses the 2015-2019 American Community Survey (ACS) 5-year Public Use Microdata Sample (PUMS) for classifying the Employment Status of Visually-impaired Graduates. The original data had about 15.95 million records; it was then filtered to contain the population being visually impaired (or reported difficulty with vision), aged 21 to 64, and had at least a Bachelor's degree. Support Vector Machines (SVMs) are a supervised machine learning algorithm used for classification and regression tasks. It identifies the best boundary known as a hyperplane that separates different classes in the data. In this study, linear, radial basis function (RBF), and sigmoid kernels were compared using One-vs-One (OvO) and One-vs-Rest (OvR). The filtered dataset contains 29,714 rows and 23 columns. From the dataset fields, WKW, and WKHP a three-class target was derived. After the target-related fields were excluded, 14 predictors remained. The data was split into training and testing data in a 80:20 split and feature scaling was fitted on the training data. Each model was evaluated with and without balanced class weights. The highest reported accuracy was 65.02% for the unweighted RBF-OvO model, closely followed by RBF-OvR at 65.00%. Although the unweighted models achieved higher accuracy, they showed very low recall for the “Less than FT/FY” class. Balanced class weights improved recognition of the less frequent class in several configurations, while reducing overall accuracy. These observations show that accuracy alone is not enough to determine the performance of this model on the dataset.
Keywords: Support Vector Machines, Employment Status Classification, SVM Kernels, One-vs-Rest, One-vs-One, Class Imbalance
Navya N, Tanushree D Joshi, Hema V, Ganeshraj R, Dr.Kanakaraju R
DOI: 10.17148/IJARCCE.2026.151151
Abstract: Preparing for an interview often involves repeated practice, feedback from peers, and advice from experienced mentors. These approaches do not always give candidates a realistic chance to practice or provide consistent feedback. The AI Mock Interview Analyzer is a web-based platform for role-specific interview practice. Users select an interview type, answer generated questions, and record their responses using a microphone and webcam. The system is designed to transcribe spoken answers and assess their relevance, language, speech characteristics, and visible facial cues. It brings these indicators together in a structured report that highlights strengths and areas for improvement. Users can also save their details and interview results for later review. By bringing question generation, response recording, analysis, and feedback into one workflow, the platform aims to make interview practice more accessible and focused.
DESIGN AND DEVELOPMENT OF ALGORITHM FOR IMAGE QUALITY ENHANCEMENT IN COMPUTED TOMOGRAPHY
Shrusti Jain N J, Srushti M, Swetha L, Vandana, Dr. Kalpana A B
DOI: 10.17148/IJARCCE.2026.151152
Abstract: Historical procurement records can support price benchmarking, but selecting a suitable reference is challenging when similar descriptions represent different brands, models, specifications, commercial units, quantities, or procurement contexts. This paper examines the historical candidate-reranking component developed for procurement price decision support in CivicEngage. For each query, an upstream retrieval stage supplies ten historical candidates. The proposed reranker combines semantic similarity, structured-attribute matching, retrieval signals, and historical price- consensus features using a LightGBM binary classifier with a nested-cross-validation-based decision gate. It either retains the original Rank-1 candidate or replaces it with a candidate from the supplied set; the returned benchmark is the selected candidate's original unit price, not a newly generated or adjusted price. On the documented benchmark of 10,560 queries and 105,600 candidate rows, the gated reranker achieved 38.99% Accuracy@10, compared with 37.57% for the original Rank-1 baseline, an improvement of 1.42 percentage points. On a clean subset of 9,417 queries whose target descriptions did not contain explicit price strings, it achieved 39.70%, compared with 38.36% for the baseline. These results show a measurable but limited improvement for the evaluated candidate pool. They do not establish general current-market price accuracy, and the result remains constrained by whether a suitable historical candidate is retrieved.
Dr. Beenarani Manoj, Sauban Tarique, Preetam N, Yogesh P
DOI: 10.17148/IJARCCE.2026.151153
Abstract: This project aims to make attendance recording easier by using face recognition to identify registered individuals automatically. The system uses YOLO11n to detect people in camera frames and a custom IoU tracker to follow their movement. InsightFace and ArcFace are used to extract facial features, while FAISS compares them with previously stored face records to find matching identities. It also checks face quality and uses liveness detection methods such as texture analysis, blink detection, motion analysis, screen-edge detection, and MiniFASNet to identify possible spoofing attempts.
The recognition results are evaluated by an AMFR decision engine before attendance is recorded in the database. Developed using Python, OpenCV, FastAPI, and Streamlit, the system also provides a dashboard for managing and viewing attendance records. A separate evaluation using six identities and 30 unseen test images achieved 86% classification accuracy using pretrained ArcFace embeddings and a newly created FAISS gallery. Although these results are promising, further testing with more individuals and under different classroom conditions is needed to assess the system's overall accuracy and reliability.
Laptop Battery Report Using Random Forest Regression Algorithm, Machine Learning
Dr. Beenarani Manoj, Hariharan N, Jeron Sam Varghese, Akash Kumar
DOI: 10.17148/IJARCCE.2026.151154
Abstract: Everyone who uses a laptop heavily has been caught out by a battery that ran flat sooner than expected. The percentage on screen only tells you what is left right now. It says nothing about how quickly that number is falling or where it will be in the next half hour. We wanted to close that gap, so we built a Machine Learning-Based Laptop Battery Drain Forecasting and Early Warning System that looks at how a laptop is being used and predicts its battery level ahead of time.
The system works in a few steps. A Python script keeps logging battery percentage, CPU usage, RAM usage, charging status and elapsed time into a CSV file. We tidied that data with Pandas and used it to train a Random Forest Regressor, which we saved with Joblib so it could be reused without retraining. A Streamlit dashboard then loads the model and puts the current battery status beside the predicted percentage. If the prediction falls below a safe level, a warning appears, giving the user a chance to plug in or save their work before the laptop shuts down. Put together, the project shows how simple data logging and a basic machine learning model can turn an ordinary battery indicator into something that helps people plan their work.
Keywords: Laptop Battery Monitoring, Battery Drain Forecasting, Machine Learning, Random Forest Regressor, Predictive Analytics, Python, Streamlit, Early Warning System
Smart Farming System Using Machine Learning for Crop Disease Detection and Precision Spraying
Payal C. Patil, Vinit S. Bari, Bhavesh D. Chaudhari, Sumit B. Khedwan, Prof. Priyanka D. Lanjewar
DOI: 10.17148/IJARCCE.2026.151155
Abstract: Crop diseases cause serious yield losses, and many farmers cannot reach an expert quickly enough to identify them. This paper presents a Smart Farming System that uses machine learning for crop disease detection and precision spraying. A farmer uploads a photograph of a crop leaf through a web interface, and a multimodal vision model classifies the leaf as healthy or diseased and reports the disease name with a confidence score. The system then produces treatment and preventive advice and, after the farmer approves, sends a command to a spraying unit that treats only the affected area. The platform also provides soil and weather information, crop management, market price trends and a chatbot to support everyday farm decisions. The proposed approach aims to reduce unnecessary pesticide use, lower cost and encourage timely action. Quantitative performance results will be reported after field testing.
ISL-HandNet: A Real-Time Indian Sign Language Recognition System Using MediaPipe Landmarks and a Random Forest Classifier
Jaithraa Madhav M, Deepak K, Prataparao Charith, Dr. G. Paavai Anand
DOI: 10.17148/IJARCCE.2026.151156
Abstract: Sign language is the primary means of communication for millions of deaf and hard-of-hearing people in India, yet very few hearing people are fluent in Indian Sign Language (ISL). Vision-based translators that classify raw RGB frames with deep convolutional networks are sensitive to background clutter, lighting changes and skin-tone variation, and usually need GPU hardware. This paper presents ISL-HandNet, a lightweight, real-time recognition pipeline for the 26-letter ISL finger-spelling alphabet that replaces pixel-level input with normalized skeletal landmarks. MediaPipe Hands extracts 21 three-dimensional hand keypoints per frame, producing a 63-dimensional feature vector. Translation invariance is obtained by expressing every landmark relative to the wrist, and scale invariance by dividing by the wrist- to-middle-finger-MCP distance. An ensemble Random Forest of 100 decision trees is trained on 17,589 valid landmark vectors extracted from 38,612 raw frames. On a held-out set of 3,226 frames the model reaches an overall accuracy of 87.10% (macro F1 of 0.86) with sub-millisecond classifier inference on a standard CPU. The system is verified with an automated test script and deployed as a live OpenCV webcam application. The results show that landmark-based feature engineering is a practical, low-cost basis for assistive ISL communication tools.
Keywords: Indian Sign Language; Hand Landmarks; MediaPipe Hands; Random Forest; Gesture Recognition; Computer Vision; Assistive Technology
BHARAT-REDRESS: An Autonomous Multimodal Ingestion and Cryptographic SLA Enforcement Architecture for Public Grievance Governance
Ketan Mohate, Vivek Maske, Prof. Panmand S. S
DOI: 10.17148/IJARCCE.2026.151157
Abstract: Citizen grievance redressal is a fundamental barometer of state administrative efficacy and democratic accountability. In linguistically heterogeneous and demographically sprawling developing democracies such as India, existing centralized portals (e.g., CPGRAMS, state CM helplines) face severe operational bottlenecks: extreme vernacular linguistic exclusion, manual subjective triage resulting in departmental misrouting, lack of automated priority quantification, and opaque, non-verifiable administrative closure without citizen consensus. This paper presents BHARAT-REDRESS, an exhaustive, decentralized, and intelligent enterprise framework designed to automate the entire citizen grievance lifecycle from intake to resolution. BHARAT-REDRESS introduces a four-tier architecture combining: (1) a Multimodal Ingestion Layer integrating national Bhashini Speech-to-Text (ASR) pipelines covering 22 official scheduled languages, computer vision-based image tamper verification, and IVR accessibility; (2) an Intelligent AI Semantic Engine utilizing a fine-tuned IndicBERT transformer coupled with zero- shot domain classifiers and a deterministic Priority Urgency Index (PUI) formulation; (3) a Consortium Blockchain Ledger orchestrated on Hyperledger Fabric featuring self-executing Service-Level Agreement (SLA) smart contracts with dynamic auto-escalation; and (4) an Asymmetric Cryptographic Resolution Engine mandating citizen-held One- Time Tokens (OCT) to prevent unilateral ticket disposal. Comprehensive empirical validation on a curated multi-tier dataset of 35,000 real-world Indian citizen grievances across 8 major language families demonstrates an automated routing classification accuracy of 93.8% (Macro F1: 0.935), an average SLA turnaround reduction of 74.2%, and transaction throughput sustained at 1,450 transactions per second (tps) under load. This end-to-end framework offers a rigorously validated, tamper-resistant, and constitutionally inclusive operational blueprint for modern digital public infrastructure.
Keywords: Digital Public Infrastructure, Grievance Redressal, Multilingual Natural Language Processing, IndicBERT, Consortium Blockchain, Hyperledger Fabric, Smart Contracts, SLA Enforcement, E-Governance.
Abstract: Deep learning models for brain tumor classification from MRI routinely report accuracies above 98% on public benchmarks. Many of these benchmarks, however, contain duplicate and near-duplicate images shared between training and test sets, which inflates reported performance. This study proposes a lightweight convolutional neural network (CNN) for four-class brain tumor classification (glioma, meningioma, pituitary tumor, no tumor) and evaluates it under a leakage- free protocol. The widely used Kaggle Brain Tumor MRI dataset (7,200 images) was first deduplicated with perceptual hashing and pixel-level verification, removing 1,994 duplicate images (27.7%); 39.5% of the original test images had a near-duplicate in the original training set. The remaining images were re-split into stratified training, validation and test sets. The proposed network combines depth wise-separable inverted residual blocks with efficient channel attention (ECA), and has only 0.60 M parameters and 0.17 G multiply–accumulate operations (MACs). On the deduplicated test set it achieves 97.3% accuracy and 97.0% macro F1-score, compared with 98.2% accuracy for the best of five standard CNN baselines (EfficientNet-B0), which have 3.7 to 224 times as many parameters. Five-fold cross-validation gives 97.2 ± 0.6% accuracy. Training the same models on the original split raised accuracy by 1.3–2.0 percentage points, quantifying the effect of leakage. To know the effect of each component, an ablation experimentation is performed and for explainability, Grad-CAM visualizations has been performed to show that predictions focus on tumor regions. The proposed model is much situated to clinical application with low resources as it can classify an MRI image in 10.1 ms on an average.
Keywords: brain tumor classification, lightweight CNN architecture, efficient channel attention, data leakage, deduplication, explainable AI
Training Data Diversity and Distribution Shift in AI-Generated Image Detection: An Empirical Study
E Niranjan, Arushi, Deesha S Chajer, Niveditha S
DOI: 10.17148/IJARCCE.2026.151159
Abstract: The proliferation of high-fidelity generative models such as Stable Diffusion, DALL-E, MidJourney, and Gemini has made automated detection of AI-generated images an increasingly important research problem. While many published detectors report high in-domain accuracy, their performance under distribution shift is often overlooked. This paper presents an empirical study of binary real-versus-AI image classifiers, examining how training data composition and architecture choice affect accuracy on images from sources not seen during training. Models trained only on public data (ResNet18 and EfficientNet-B3) reached 53.33% and 61.67% on a 60-image held-out set of images from web-based generators and personal photographs, whereas a ResNet18 trained with an additional 360 self-collected images reached 75.00%, rising to 81.67% after threshold calibration. On unseen Tiny-GenImage samples (n=800) the three models scored between 72.12% and 74.75%, and the data-augmented model scored lower than the public-only ResNet18 on a 2,000- image slice (68.85% vs. 72.60%), indicating a trade-off between the added sources and the original public distribution. Because the two architectures have near-identical parameter counts (11.18M and 10.70M) and the runs differ in training volume and schedule, we draw no conclusion about model capacity. The results suggest that adding even a small amount of data from the target sources can matter more than the choice between these two architectures.
Keywords: AI-generated image detection, deep learning, convolutional neural networks, dataset diversity, transfer learning.
AI-Based Automatic Modulation Classification (AMC) Using a Convolutional Neural Network (CNN)
Nalini G, Prakashkumar B Tole, Kapileshwara N Katti, Mahesh CS
DOI: 10.17148/IJARCCE.2026.151161
Abstract: Automatic modulation classification (AMC) identifies the modulation format of a received signal without help from the transmitter and underpins spectrum monitoring and cognitive radio. This paper presents a reproducible, end-to- end AMC framework that is built on a synthetic signal generator instead of a fixed benchmark file. Six schemes (BPSK, QPSK, 16-QAM, AM, FM and FSK) are synthesised, passed through AWGN, phase-noise and Rayleigh-fading channels at SNRs from −10 dB to 30 dB, and converted to 2×128 baseband I/Q frames by Hilbert-transform preprocessing. Two classifiers are studied: an interpretable baseline that uses six envelope and instantaneous-frequency statistics with Gaussian class models, and a convolutional neural network (CNN) whose layer structure follows a published Gaussian- noise-regularised design. The baseline reaches overall accuracies of 61.1%, 69.2% and 59.4% under AWGN, phase noise and Rayleigh fading respectively, and roughly 80% or more once the SNR is 10 dB or higher; BPSK/QPSK confusion is its dominant failure mode. The CNN reaches 89.5% overall test accuracy on a mixed-channel test set. An interactive dashboard exposes every stage and produces accuracy-versus-SNR curves, confusion matrices and training curves.
FUTURE-AWARE JOB MARKET DEMAND FORECASTING AND RESUME MATCHING USING MACHINE LEARNING
Shri Nivetha R, Kavya S Prasad, Jeyaathri V, Dr. G. Paavai Anand
DOI: 10.17148/IJARCCE.2026.151162
Abstract: This paper presents a future-aware job market analytics and resume matching system that combines monthly job-posting demand forecasting with skill-trend analysis and candidate skill-coverage assessment. A global Random Forest regression model is trained using lagged demand, rolling statistics, growth indicators, seasonal features, forecast horizon, series type and job-domain information. A chronological hold-out design is used to evaluate twelve-month- ahead predictions against naïve and seasonal-naïve baselines. Forecasted skill demand is then combined with recent role- specific skill relevance to construct a weighted target skill profile. Resumes are parsed for recognised skills and assessed using a transparent weighted coverage score, while TF-IDF cosine similarity is reported separately as a supplementary text-alignment measure. On the supplied artificial dataset, the model achieved MAE 3.07, RMSE 4.31 and R² 0.908, outperforming the pooled baselines. These findings demonstrate the feasibility of integrating forecasting and explainable skill-gap analysis, but do not establish real labour-market performance.
Keywords: Machine Learning, Random Forest Regression, Job Market Forecasting, Skill Trend Analysis, Resume Parsing, TF-IDF, Skill Gap Analysis.
Tamil Nadu District-Level Next-Day Rainfall Prediction Using Explainable Machine Learning and a Two-Stage Hurdle Architecture
A Soujanya Lakshmi, Ruthiksha P, Yuvaraj P
DOI: 10.17148/IJARCCE.2026.151163
Abstract: This paper presents an end-to-end machine learning framework for forecasting next-day rainfall depth, in millimetres, and rain occurrence across all 38 administrative districts of Tamil Nadu, India. Because precipitation data are strongly zero-inflated (roughly 80% dry days) and right-skewed, a Two-Stage Hurdle Architecture is proposed: a CatBoost classifier gate estimates whether rain will occur (≥ 0.1 mm), and a CatBoost regressor estimates the rainfall depth. Using the TNWeather24 dataset (2019–2024), raw records are aggregated to district-daily values, and 27 leakage- safe features built only from observations up to the 23:59 IST cutoff are used to train the models. Evaluation is strictly chronological (training 2019–2022, validation 2023, unseen test 2024 with 13,870 records) and compares the hurdle model with Random Forest, XGBoost, a single-stage CatBoost regressor, and a historical-climatology baseline. On the 2024 test set the hurdle model reaches a rain-occurrence accuracy of 82.14%, a ROC-AUC of 90.56%, a rain-day recall of 81.80%, and a depth MAE of 0.1393 mm, a 34.2% reduction relative to the baseline. SHAP TreeExplainer provides global and local explanations, and a six-page Streamlit dashboard and 14 passing automated tests support use and verification of the framework.
Abstract: Investor news filters usually match company names, tickers or keywords, or compare embeddings of a query and a headline that were computed separately. These methods handle company announcements well but miss a different kind of risk: a problem at a supplier, a raw-material source or a transport route that affects a company without the headline ever naming it. We call these second-order shocks. FinNews detects them by describing each monitored company with a three-part text profile (identity, sector and upstream inputs) and scoring every headline–profile pair with a cross-encoder, which reads both texts together instead of relying on name overlap. The scorer is a DeBERTa-v3-small model initialized from a natural language inference (NLI) checkpoint and fine-tuned on 66,606 pairs covering 55 companies in 10 sectors, split chronologically into training, validation and holdout sets. On the 6,661-pair holdout set at a threshold of 0.40, the model reaches 74.06% accuracy, 71.67% precision and 79.16% recall; it retains 77.78% of second-order shocks and rejects 69.01% of irrelevant headlines. A 12-layer MiniLM passage-ranking baseline, scored on an earlier holdout without enriched profiles, retains 47.60% of second-order shocks. Because the backbone and the profile format changed together between these runs, the gain cannot yet be assigned to either. On a small supplementary live set collected on 10–12
shocks (83.33%). We also report an evaluation error in which a second sigmoid applied to already-normalized scores makes every pair pass the threshold, and we describe a CPU-only deployment that scores a pair in 46–66 ms.
The Connectome Lab: Connectome-Constrained Neural Chess
T Milan Prabhu, Gokul J, Sudarshan J, Dr. Beenarani Manoj
DOI: 10.17148/IJARCCE.2026.151166
Abstract: We present The Connectome Lab, an interactive system that connects browser chess positions to connectome-constrained neural models through an authenticated model API. The interface combines a fixed, untrained leaky integrate-and-fire model with MFly-1.0, a larger chess-trained recurrent rate model, and displays action selection, model identity, and recorded activity. The deployed trained path uses a pretrained checkpoint; a separate project training workflow creates and exports project checkpoints. A 2,000-neuron CPU experiment completed imitation learning and short self-play, reaching checkpoint step 68. Its final logged imitation check at step 60 produced 1.56% top-one and 3.12% top-three move agreement on 64 held-out positions, with value mean squared error of 1.0007. The exported checkpoint recorded four draws against a random opponent and two draws and two losses against a material opponent in eight bounded games. Saved API observations and six browser fixtures verify connected decisions and action handling. The work demonstrates an integrated experimental application and a training pipeline. The brief experiment and unmatched model configurations do not establish competitive chess strength or an advantage from biological wiring.
Keywords: connectome-constrained networks, neural chess, leaky integrate-and-fire, recurrent neural networks, imitation learning, self-play, model API, recorded activity.
Abstract: The Intelligent Health and Fitness Assistant is a smart fitness application designed to help users monitor and improve their physical health through technology. The system uses Python, Streamlit, OpenCV, MediaPipe, and Artificial Intelligence to track body movements and analyze exercise activities using a camera.
The application detects the user's body posture through pose estimation and provides feedback on exercise movements, repetitions, and workout performance. It can also calculate Body Mass Index (BMI) and provide basic fitness-related guidance based on user information. The system uses text-to-speech (gTTS) to provide voice feedback, making the interaction more convenient for users.
The main objective of this project is to provide an accessible, interactive, and intelligent fitness assistant that can be used without expensive fitness equipment. By combining computer vision, AI, and a user-friendly interface, the system helps users perform exercises more effectively and maintain a regular fitness routine.
Keywords: Intelligent Health and Fitness Assistant, Machine Learning Algorithms, Human Activity Recognition, Pose Estimation (MediaPipe), Exercise Posture Correction, Personalized Workout Recommendations, Health Risk Assessment, Fitness Tracking, Real-Time Activity Monitoring, Web Application.
DESIGN AND DEVELOPMENT OF A LoRa-BASED SMART SAFETY HELMET FOR REAL-TIME MONITORING OF UNDERGROUND MINE WORKERS
Dr. Vijayalakshmi D, Chandra Shekar NM, Chetan Kumar H, Ambaresha, Harsha H
DOI: 10.17148/IJARCCE.2026.151168
Abstract: Underground mine workers operate in environments where hazardous atmospheric conditions and worker- related emergencies require timely monitoring. This paper describes the design and prototype implementation of a wearable smart safety helmet integrating environmental and selected physiological sensing with wireless data communication. An ESP32 transmitter acquires information from MQ-4, MQ-135, DHT11, MAX30100 and a GPS receiver. An SX1278-based LoRa module transmits the data to an Arduino Nano receiver, and a laptop dashboard displays the received fields. A buzzer provides a local alert output according to configured firmware conditions. The assembled prototype was demonstrated with a dashboard showing sample sensor readings and helmet status. The displayed readings are preliminary observations, and further sensor calibration, repeated testing, underground radio characterization, and mine-safety certification are required to establish the system's reliability and suitability for practical mining applications.
Abstract: Diagnostic errors and delays in emergency triage remain problems in healthcare. Artificial intelligence can help doctors make decisions; however many systems out there don't explain why they make the predictions they do. This paper introduces MedAgentix AI, a multi-agent clinical decision support system. MedAgentix AI is designed to help with diagnosis, disease risk assessment and emergency triage. The system uses machine learning, deep learning and natural language processing. These tools work together through agents in a single workflow. MedAgentix AI takes symptoms from what patients say makes predictions checks for possible risks and looks through medical documents to find important information. It also checks for emergency warning signs to spot cases. Retrieval-Augmented Generation gives doctors the medical information to answer questions and handle cases where the system is not sure. The system creates reports for doctors and for patients. The data, behind MedAgentix AI includes 246,945 cases. These cover 773 conditions and 377 standard symptoms. The systems multi-class diagnostic model works with 721 types of diseases. The findings show that multi-agent intelligence and explainable artificial intelligence can help make clinical decisions more transparent. More testing is needed to make sure MedAgentix AI is reliable and can be used in hospitals.
Keywords: Artificial Intelligence in Healthcare, Clinical Decision Support Systems, Multi-Agent Systems, Explainable Artificial Intelligence, Differential Diagnosis, Emergency Triage, Natural Language Processing, Machine Learning.
Abstract: Cloud storage has become an essential backbone for personal and organizational data management in the modern digital era. However, existing commercial cloud architectures such as Amazon Web Services (AWS), Google Cloud, Dropbox, and Microsoft Azure rely heavily on third-party hosting, where external technology corporations control the physical servers and security infrastructure. This dependence introduces significant limitations for users, including a lack of data sovereignty, recurring subscription and hidden operational costs, and vulnerability to third-party service outages. Furthermore, while standard hybrid cryptography models combine the Advanced Encryption Standard (AES) for rapid payload encryption with Rivest-Shamir-Adleman (RSA) for secure key transport, they suffer from a critical blind spot: the absence of data integrity verification. If an attacker intercepts an encrypted file and flips random bits, or if the file becomes corrupted over an unstable network connection, a standard AES and RSA setup provides no warning and decrypts a corrupted or maliciously altered file silently.
Simple Sovereign Cloud Storage (SSCS) is a fully sovereign, globally accessible cloud storage environment hosted entirely on local hardware to eliminate reliance on third-party vendors. The system exposes a local hard drive storage medium as a distributed file system using Network File Share (NFS) on a Unix-based operating system, guarded by a customized firewall (such as firewalld or ufw) to restrict unauthorized and malicious network traffic. Client devices interact with the cloud server through a command line interface (CLI) that performs encryption and decryption. It uses AES to encrypt the file data, and Secure Hash Algorithm 256-bit (SHA-256) generates and verifies cryptographic digests to guarantee end-to-end data integrity.
Keywords: Sovereign Cloud Storage, Advanced Encryption Standard (AES), Data Integrity Verification, Network File System (NFS)
Abstract: Large language models can produce answers that sound right but are not actually supported by the text they were given. This work builds a simple, fully CPU-based check for that problem. Each answer is turned into five numbers: the share of hedge words, the average token likelihood under a small local model, how similar the answer is to the evidence, how similar it is to the question, and an entailment probability. Those five numbers then go into ordinary classifiers (logistic regression, random forest, XGBoost, a tiny neural net) and also into two quantum-style classifiers that run in simulation: a five-qubit variational circuit and a quantum-kernel SVM. We took the first sixty questions from the HaluEval QA set. That gives one hundred and twenty answers—one correct and one hallucinated for every question. Splits keep whole questions together so nothing leaks. On the sixty-answer training slice that both families see, the classical models reach an AUROC of about 0.975. The quantum-kernel SVM gets 0.90 and the variational circuit only 0.53. The test set itself is small (eighteen answers). A later check of the exact quantum state also showed that our particular circuit almost completely ignores its fifth input. The code is reproducible, models reload cleanly, and a small local interface lets anyone try new examples. The pipeline works, but the data set is tiny, calibration is uncertain, and the variational circuit has a clear design flaw, so we make no claim of quantum advantage.
Abstract: In Bagalkot, many farmers travel one to two kilometers at night to switch on their bore well pump. Sometimes they forget to switch off the pump. Sometimes the pump keeps running even when rain starts. This wastes water and electricity. To solve this daily problem, our team made a small model called JALA RAKSHAKA. In Kannada, Jala means water and Rakshaka means protector. The model has two sensors. One sensor checks whether the soil is dry. Another sensor checks whether rain has started. An Arduino Uno board reads both sensors and decides whether the pump should run or stop. We collected black soil from a field near Bagalkot and tested the model. When the soil was dry and there was no rain, the pump started automatically. When the soil became wet or rain drops fell on the rain sensor, the pump stopped automatically. The model works without internet and without a mobile phone. The total cost was eight hundred and seventy rupees only. This model got Second Prize in Project Exhibition at our college.
Sign Language Translation Using Machine Learning and Artificial Intelligence
Srinidhi K S, Nayan Kumar J S, Nikhil S, Nikhil M, Mohith S Gowda
DOI: 10.17148/IJARCCE.2026.151178
Abstract: Communication between sign language speakers and non-sign language speakers relies on human interpretation and/or sensors, both restricting frequent usage. In this paper, we describe the design of a webcam-based application capable of recognizing signs in sign language through a Convolutional Neural Network (CNN) and showing them as text and also speech as a planned feature. The architecture consists of dataset formatting and labeling, pre- processing of images, training of CNNs with the help of the Adam optimizer and categorical cross-entropy loss function and recognition process with OpenCV, TensorFlow and Keras running on regular Windows laptop. Signs or gestures are to be put together into simple sentences with planned features such as suggesting the next character and conversion into speech. The project is in an intermediate phase where 5 Kaggle datasets are examined for appropriateness and around half of the training process is complete. There are no performance measurements yet: only the evaluation procedure with accuracy, precision, recall, F1 score and confusion matrices is stated and the other limitations are discussed.
Keywords: Sign Language Recognition, Computer Vision, Convolutional Neural Network, Deep Learning, Gesture Recognition, Text to Speech.
A Review of Control Architectures, Perception Methods, and Validation Approaches of Adaptive Headlight Control System
Aparna P. Goswami, Dr. C. Srinivasa Murthy, Rachana K. Chavan
DOI: 10.17148/IJARCCE.2026.151179
Abstract: Conventional automotive headlamps provide a largely fixed illumination pattern and often depend on drivers to switch manually between high and low beams. This creates two related safety challenges: insufficient illumination around bends or uneven roads, and glare to other road users when high beams are not reduced in time. This review synthesizes ten supplied papers on adaptive front-lighting and intelligent headlight control, covering low- cost light-sensor circuits, steering- and inertial-sensor-based beam positioning, camera-based road-curvature estimation, machine-learning object detection, camera–radar and camera–LiDAR fusion, vehicle-to-vehicle (V2V) heading exchange, and vehicle-in-the-loop (ViL) verification. The studies indicate that no single sensing or control approach addresses all operating conditions. Threshold-based circuits are inexpensive and easy to implement but can confuse different light sources; vehicle-dynamics sensors support beam steering but do not directly identify hazards; and vision or fused-perception systems offer richer context while introducing computation, calibration, synchronization, and validation requirements. Experimental work on dipped-beam visibility further shows that beam shape and road class influence useful illumination. A key conclusion is that algorithmic capability must be paired with measurable performance criteria and repeatable testing. The review proposes a layered architecture combining perception, confidence-aware fusion, constrained beam control, safe fallback behavior, and quantitative verification. Because the supplied papers differ in prototype, scenario, and evaluation method, their results are compared qualitatively rather than pooled statistically.