A production-grade computer vision platform automating railway spring health classification using YOLOv8, real-time WebSocket pipelines, and Groq Vision LLMs.
Fine-tuned YOLOv8 on 10,000+ annotated images achieving 98.27% mAP@50.
Engineered real-time 30 FPS WebSocket video pipeline with BoT-SORT tracking and IoU deduplication.
Integrated Groq Vision API (Llama-3.2-11b) for spring condition classification (Healthy/Degraded/Critical).
Delivered a React.js dashboard with Supabase audit trails for predictive maintenance.
Multimodal Medical Report Captioning
Vision-Language Diagnosis Captioning
A deep learning pipeline that generates accurate clinical captions for medical reports using a vision-language framework with retrieval-augmented generation.
Designed a ViT encoder + GPT-2 decoder pipeline for clinical image feature extraction and text generation.
Integrated RAG over a medical corpus, improving caption factual accuracy by 30%.
Used OpenAI API for clinical relevance refinement and automated evaluation.
RAG Customer Support Chatbot
Stateful Agentic Customer Support Chatbot
An advanced customer support agent orchestration system built with LangGraph, utilizing hybrid dense-sparse retrieval and structured tools.
Built a state-of-the-art LLM chatbot with hybrid dense + sparse retrieval, reducing irrelevant responses by 40%.
Used LangGraph for stateful multi-step agent orchestration with tool-calling and structured decision branching.
Legal Case Outcome Prediction
NLP Classification & Knowledge Graphs for Indian Case Precedents
An NLP document classification model coupled with a large knowledge graph designed to recommend relevant legal precedents and predict case outcomes.
Fine-tuned InLegalBERT for precise Indian legal document classification.
Integrated a 50,000+ node Neo4j knowledge graph with Cypher queries for precedent-based recommendations.
Like this project
Posted Jun 3, 2026
Automating railway inspection and enhancing document intelligence using AI.