Interactive Multimodal RAG Chatbot with Automated PDF Reporting for Model Training Analysis
Designed and developed an interactive multimodal RAG chatbot using Gemma-3 (12B) to analyze model training reports and logs. Implemented a RAG pipeline over PDFs using MiniLM text embeddings and SigLIP image embeddings for visual-aware retrieval; extracted PDF images and generated contextual captions with a multimodal LLM. Built automated PDF reporting that consolidates YAML configurations, training logs, images, prompts, and LLM analysis, along with a Gradio UI for contextual Q&A, follow-up suggestions, and downloadable reports. Used Python, PyTorch, Hugging Face Transformers, CUDA, LangChain, ChromaDB, SigLIP, and Gradio.
AI-Powered Multi-Agent Deep Research System
Designed and developed a LangGraph-based multi-agent research system with agents for query clarification, research planning, web search, evidence synthesis, and report generation. Implemented hierarchical Supervisor, Research, Scoping, and Compression agents; integrated OpenRouter and Tavily Search API; built a FastAPI backend for REST APIs; and extended the platform with MCP support for local files, PDFs, and GitHub repositories. Used Python, LangGraph, LangChain, FastAPI, OpenRouter, Tavily Search, MCP, and Pydantic.
Real-Time Pothole Detection & Segmentation Pipeline
Developed a Mask R-CNN (ResNet-50 + FPN) instance-segmentation system for pothole detection. Built preprocessing, model-training, and evaluation pipelines using MMDetection; achieved 0.72 mAP; addressed varying lighting, road textures, and occlusions; and accelerated inference with TensorRT FP16 to approximately 9.58 ms latency. Used Python, PyTorch, MMDetection, OpenCV, Docker, and TensorRT.
Video-Based OCR Pipeline with Confidence Stabilization
Developed an end-to-end OCR system for extracting structured text from videos using frame sampling, ROI extraction, and preprocessing. Implemented confidence normalization, temporal stabilization, numeric validation, noise removal, and error-correction rules, and evaluated OCR performance across diverse video conditions. Used Python, OpenCV, Paddle OCR, TrOCR, NumPy, and Pandas.