Architected a GraphRAG pipeline with LangChain, Neo4j, Cypher, FastAPI, and LangGraph, combining entity linking, subgraph expansion, hybrid retrieval, and reranking for citation-grounded response generation over a legal knowledge graph. Established evaluation and observability with RAGAS, tracking faithfulness, relevance, F1 score, and retrieval quality to harden response reliability on ambiguous queries. Optimised local inference using GGUF/4-bit quantisation, KV cache, FlashAttention, and speculative decoding, achieving faster generation with reduced memory footprint.

Developed scalable FastAPI backend with microservice architecture, dependency injection, and layered design for maintainability and independent service scaling. Built AI-powered voice assistant using ElevenLabs Client SDK to guide users, teach vocabulary, and evaluate learning through AI-generated quizzes. Optimised performance through Redis caching, rate limiting, memoisation, and database query tuning, reducing response latency across high-traffic endpoints. Containerised with Docker and deployed via CI/CD pipelines on GitHub Actions to VPS.

Built scalable backend with Spring Boot applying layered and domain-driven design for modularity and separation of concerns. Architected stateless JWT authentication with RBAC, enabling horizontally scalable session management and secure access control. Optimised MongoDB persistence layer with indexing, connection pooling, and pagination to support high-throughput ingestion and retrieval. Engineered real-time notifications via Server-Sent Events (SSE), decoupling read-heavy streams from core paths to improve overall responsiveness.

Built AI voice agents with intent routing, tool-calling, and stateful dialogue for fully automated reservations, ordering, and customer support over phone. Integrated Twilio and Vapi for speech-to-speech pipelines, implementing event-driven call routing, spam filtering, and human escalation with async callback automation. Developed OCR-based menu ingestion using vision-language models for structured extraction across multi-modal restaurant data pipelines.

DubAI is a web application that leverages advanced AI technologies to dub Youtube videos in over 20 languages. The platform utilizes state-of-the-art models for automated transcription and translation, ensuring high-quality audio separation and dubbing. Users can easily upload their youtube video link, select the desired language, and receive a dubbed version that seamlessly integrates with the original video.

The Shape Drawer project is a graphical application that allows users to draw, manipulate, and interact with various shapes on a canvas. This project is built using C# and leverages the SplashKitSDK library for graphical rendering and event handling. The application supports drawing shapes, with features such as shape selection, color changes, and background customization.

The Bus E-Ticketing System allows passengers to reserve seats online, while administrators can manage fleet schedules and seat availability in real-time. With secure authentication, real-time booking, and integrated payment processing, the system enhances the public transportation experience by offering a seamless and digitalized ticketing process.

Awesome Weather App allows users to instantly check the weather of any city without requiring authentication. It features a visually appealing interface with interactive GIFs that change dynamically based on weather conditions. Alongside temperature updates, users can view humidity, wind speed, and other weather details.

This project focuses on breast cancer detection using machine learning techniques. We utilized the Breast Cancer Wisconsin dataset, which contains diagnostic data for breast cancer patients. The project involved data preprocessing, feature selection, and model training using various algorithms to achieve accurate predictions.

This project focuses on detecting signboards using YOLOv5 model. We annotated images using CVAT, prepared a structured dataset, and trained the YOLOv5 model for signboard detection. After training, we evaluated the model's performance and applied it to predict signboards in new images.

WatchDog is a surveillance robot designed to monitor and secure indoor spaces. It features a live video streaming function, enabling users to remotely view the robot's surroundings with access of multiple sensory data.

Showcased at TechSpectra, this pick-and-place robot is designed to grasp, hold, and release objects with precision. Built on an Arduino framework, it integrates two servo motors for seamless arm and claw operation. This ensures flexibility and efficiency in various tasks requiring delicate handling.