I work across production-grade RAG, multi-agentic systems, and deep learning models for various imaging probelems.
Introduces a scene-graph-first reasoning pipeline that makes VLM answers more structured by separating object relationships from higher-level chain-of-thought reasoning.
Targets tool-tissue-action triplet recognition in surgical video by using temporal weak supervision, reducing dependence on dense frame-level labels.
Builds a medical-imaging classifier for TB bacilli using weighted color and morphology channels, improving sensitivity on noisy cytopathology smears.
Built a CrewAI-based multi-agent pipeline for LinkedIn publishing: AI-news scraping, NER topic tagging, LLM drafting, activity logging, and scheduled content operations for 5+ posts per week.
Also shipped an OCR extraction pipeline with PaddleOCR and LayoutLMv3 for noisy PAN and tax documents, using regex normalization and confidence thresholds to clean field-level outputs.
Worked on cytopathology image analysis for acid-fast bacilli detection, designing a color-morphology feature fusion model around illumination, saturation, hue, and Sobel edge channels.
This line of work became the ICVGIP paper on TB bacilli detection, with channel attention, reconstruction-loss weighting, focal loss for false positives, and Grad-CAM interpretability.
Production policy-QA RAG system that strips contradicted claims at inference using DeBERTa-v3 NLI verification, hybrid Pinecone + FTS5 retrieval, RRF fusion, reranking, confidence gating, and Redis session memory.
LangGraph system with Planner, Researcher, Verifier, Analyst, and Writer agents; it routes across open and closed LLMs, stores memory in Qdrant, grounds outputs in citations, and evaluates quality with RAGAS.
Multimodal price-prediction system fusing satellite imagery through EfficientNet-V2-S, FPN-Lite, and CBAM with tabular features through an MLP, then combining both streams with late fusion.
Multimodal extraction pipeline for product attributes such as dimensions, quantities, and units from noisy listings and images, combining OCR cleanup with entity normalization for leaderboard-style evaluation.
A focused stack for building systems that can be evaluated, deployed, and debugged.
LangGraph, LangChain, CrewAI, LiteLLM, OpenAI API, vLLM, Groq, mem0.
FAISS, LanceDB, Qdrant, Pinecone, SQLite FTS5, RAGAS, LLM-as-judge harnesses.
PyTorch, Hugging Face Transformers, scikit-learn, OpenCV, spaCy, ONNX Runtime.
Docker, FastAPI, AWS, GCP, MLflow, W&B, Celery, Redis, Prometheus, MySQL, MongoDB.
Academic base and core training.
B.Tech. in Biosciences and Bioengineering, CGPA 8.27. Coursework and projects bridge biological systems, computer vision, and production ML engineering.