As a Biomedical Engineer by foundation, I leverage my domain expertise to develop advanced AI solutions. I specialize in designing Agentic Workflows, Hybrid RAG systems, and production-ready distributed backend services. I am passionate about building end-to-end AI pipelines that go beyond prototypes — from architecture decisions to deployment. Currently exploring modern RAG and CAG (Cache-Augmented Generation) architectures.
-
Paperwise — Distributed Hybrid RAG System (Active Development)
- Designed and built an end-to-end hybrid RAG system enabling natural language queries over real-time arXiv paper ingestion.
- Implemented a dual-collection Qdrant pipeline separating abstract-level semantic selection from full-text chunk retrieval — improving mean chunk relevance score from 0.421 to 0.610 across prototypes.
- Engineered LLM-driven query analysis to extract domain-optimized search terms and rank up to 80 candidate documents by embedding similarity.
- Evolved architecture from single-process CLI to a distributed microservice system (FastAPI, RabbitMQ worker pool, Redis cache) supporting concurrent multi-user sessions with token-by-token SSE streaming.
- Developed across 3 self-contained prototypes with documented architecture comparisons; LLM judge evaluation improved from 4.4/5 to 5.0/5 between Prototype 1 and Prototype 2.
-
Clinical Insight Agent (Hybrid RAG)
- Architected an autonomous agent to query ClinicalTrials.gov, using LangGraph for intelligent routing between SQL and ChromaDB.
- Developed a real-time ETL pipeline, strictly grounding LLM responses in source records to prevent hallucinations.
- Deployed an asynchronous FastAPI backend with interactive Streamlit UI.
-
End-to-End Prediction Pipeline with MLOps
- Built a production-ready ML forecasting pipeline using XGBoost and SOTA algorithms.
- Accelerated feature engineering using RAPIDS (cuDF) and tracked experiments with MLflow.
-
Brain Tumor Segmentation & Volumetric Analysis
- Developed a deep learning pipeline to automate Glioblastoma tumor segmentation from raw MRI (DICOM) data using TensorFlow and OpenCV.
- Calculated precise tumor volumes utilizing DICOM voxel spacing metadata.
