Publications
A Physics-Informed Computer Vision Framework for Firefighting Assistance
IEEE 9th International Conference on Information and Communication Technology (CICT), Chennai, India
2025
- Engineered a PINN-based fire spread predictor embedding heat transfer PDEs into the loss function alongside YOLOv8 detection, achieving 98.12% accuracy and 0.85 IoU with Kalman filter tracking (6-DOF state, IOU association) reducing false alarms by 94%.
- Integrated multi-physics modeling combining flame temperature estimation (525–3000°C), fuel load quantification across 26 object classes, and dynamic threshold adaptation achieving 85% prediction accuracy with real-time trajectory forecasting.
- Built a grid-based fire spread analysis engine with directional prediction vectors, burn-time estimation, and environmental risk scoring for emergency response decision-making.
Projects
Backend & Infra
Transaction Verification and Monitoring System
PythonFastAPIWebSockets
- Built a blockchain transaction verification system using FastAPI, combining Merkle proof generation and multi-signature verification with a unified off-chain and on-chain verification flow.
- Designed and deployed Solidity smart contracts for an on-chain verification registry, implementing batch verification to reduce gas costs versus per-transaction validation.
- Implemented real-time blockchain monitoring using WebSocket-based event subscriptions and async Python, enabling low-latency detection and processing of on-chain activity.
- Developed a cryptographic verification pipeline using keccak256-based Merkle trees with sorted-pair optimization and efficient proof construction for transaction inclusion guarantees.
OCR Enabled Bills Stock Monitoring
React NativeTesseractFirebaseFirestore
- Built a React Native bill management app with OCR using Tesseract to extract text from camera or gallery images.
- Implemented Firebase Authentication with email verification and Firestore for user data storage.
- Developed CRUD operations for bills with image capture, text extraction, and full-screen image viewing.
- Integrated React Navigation with protected routes and conditional rendering based on authentication state.
- Added camera and gallery permissions handling with image picker integration for bill capture workflows.
Blogging App
Node.jsExpress.jsMongoDBEJSJWTMulter
- Built a full-stack blogging platform using Node.js, Express.js, and MongoDB with JWT-based authentication, implementing secure password hashing via HMAC-SHA256 with salt, role-based access control (USER/ADMIN), and cookie-based session management for user authentication and authorization.
- Developed a content management system with Multer for image uploads, MongoDB population for relational data queries, and server-side rendering using EJS templates with partials, enabling users to create, view, and manage blog posts with cover images.
- Implemented a real-time commenting system with user associations and timestamps, utilizing Mongoose schema relationships and population to link comments to blogs and authors, creating an interactive user engagement feature.
- Architected a production-ready application with environment variable configuration using dotenv, dynamic port assignment for cloud deployment, static file serving middleware, and MVC pattern separation across routes, controllers, models, and authentication services.
URL Shortener
Node.jsExpress.jsMongoDBEJSJWT
- Developed a full-stack URL shortener using Node.js, Express.js, and MongoDB, implementing RESTful APIs with cookie-based JWT authentication and role-based authorization (NORMAL/ADMIN) to secure user-specific URL management.
- Built analytics tracking with MongoDB aggregation, capturing visit timestamps and click counts per shortened URL, enabling users to monitor link performance in real time.
- Implemented server-side rendering with EJS templates and Express middleware, creating a responsive web interface for URL generation, user authentication, and dashboard views.
- Architected a scalable MVC application with nanoid-based unique ID generation, MongoDB schema design with referential integrity, and middleware-based authentication flow, supporting user registration, login, and protected route access.
Blockchain
Smart Contract for CLAMM Tick collector
SolidityEVM
- Built a unified tick data collection system supporting singleton CLAMM protocols such as Uniswap V4 and PancakeSwap Infinity, abstracting pool state, slot0, tick bitmaps, and liquidity reads behind a common interface.
- Implemented gas-optimized bitmap scanning and word-level iteration to efficiently discover initialized ticks across entire pools and in price-local regions around the current tick.
- Engineered direction-aware, gas-bounded tick traversal with precise handling of tick spacing, compressed ticks, word boundaries, and current-tick edge cases.
- Applied Q96 fixed-point arithmetic, low-level bitwise math, and selective inline assembly to accurately profile liquidity for routing, arbitrage, and execution simulation use cases.
Cross-Chain ERC20 Bridge
SolidityNode.jsHardhatEthers.js
- Implemented a bidirectional ERC20 bridge using lock-mint and burn-release semantics, coordinating state across two chains via on-chain events, off-chain message construction, and destination-chain execution.
- Designed a validator-signed cross-chain authorization scheme using ECDSA with chainId + contract-address domain separation and monotonic nonces to prevent replay, cross-chain, and cross-contract signature reuse.
- Built an event-driven relayer pipeline that subscribes to Solidity events, deterministically reconstructs signed payloads, and submits authenticated transactions using role-isolated validator and relayer keys.
- Developed end-to-end Hardhat tests without mocks simulating the full bridge lifecycle, validating signature recovery, nonce enforcement, role permissions, and balance conservation across chains.
Transaction Verification and Monitoring System
PythonFastAPIWebSockets
- Built a blockchain transaction verification system using FastAPI, combining Merkle proof generation and multi-signature verification with a unified off-chain and on-chain verification flow.
- Designed and deployed Solidity smart contracts for an on-chain verification registry, implementing batch verification to reduce gas costs versus per-transaction validation.
- Implemented real-time blockchain monitoring using WebSocket-based event subscriptions and async Python, enabling low-latency detection and processing of on-chain activity.
- Developed a cryptographic verification pipeline using keccak256-based Merkle trees with sorted-pair optimization and efficient proof construction for transaction inclusion guarantees.
AI / ML
Physics-Informed Neural Network Fire Forecasting System
PythonPINNYOLOv8OpenCV
- Engineered a PINN-based fire spread predictor embedding heat transfer physics with YOLOv8 detection achieving 98.12% accuracy and 0.85 IoU.
- Implemented advanced temporal tracking using Kalman filters with 6-DOF state estimation and IOU-based object association, reducing false alarms by 94% through multi-modal filtering.
- Developed dynamic threshold adaptation system calculating fire-size dependent success thresholds, achieving 85% prediction accuracy with real-time trajectory forecasting.
- Integrated multi-physics modeling combining flame temperature estimation (525-3000°C range), fuel load quantification across 26 object classes, and environmental risk scoring for hazard assessment.
- Built comprehensive fire spread analysis engine with grid-based tracking, directional prediction vectors, and burn-time estimation supporting emergency response decision-making systems.
Multi-Class Endoscopic Disease Classification System
PythonTensorFlowCNN
- Architected hierarchical binary CNN ensemble with one-vs-all classification strategy, achieving 92.3% accuracy and 0.89 AUC across 10 pathological classes using stratified Bootstrap resampling with 500 samples per class.
- Engineered adaptive optimization framework with ReduceLROnPlateau scheduler (factor=0.5, patience=3), early stopping regularization, and dynamic learning rate decay from 1e-3 to 1e-6 for optimal convergence.
- Implemented multi-scale data augmentation pipeline with stochastic transformations (horizontal/vertical flips, 0.2 rotation variance), 224x224 pixel normalization, and automated class-balanced dataset synthesis.
- Developed maximum likelihood ensemble classifier using sigmoid activation confidence scoring across independent binary models, enabling uncertainty quantification and robust pathological feature extraction.
Multi-Modal Vehicle Tracking and Taillight Detection System
PythonYOLOv8
- Architected multi-model ensemble system integrating YOLOv8n, custom brake light, and taillight detection models with real-time object tracking achieving 30 FPS processing on 1080p video streams.
- Implemented trajectory-based motion analysis with 30-frame sliding window tracking, computing Euclidean distance metrics and normalized deviation thresholds for anomalous driving behavior detection.
- Developed adaptive visual tracking pipeline with dynamic color coding based on deviation analysis (0.05 threshold), enabling real-time identification of erratic vehicle movements and lane departures.
- Engineered synchronized multi-detection framework processing vehicle bounding boxes, brake light activation states, and taillight visibility with frame-level temporal correlation for comprehensive traffic analysis.