Darshan Gowda M

RAG · Qdrant · Cohere command-r-plus · FastAPI · PyMuPDF · TF-IDF · RapidFuzz

End-to-end RAG pipeline for FDA 510(k) predicate discovery across a corpus of cleared medical device PDFs. Extracted Intended Use and Technical Characteristics at temperature=0.0 using a two-layer similarity system — dense vector cosine search for Intended Use matching and TF-IDF sentence-level matching for technical characteristics. Auto-generated compliant 510(k) Summary Reports via LLM.

Azure OpenAI GPT-4o · BM25Plus · langextract · PyMuPDF

Retrieval-augmented pipeline without embedding vectors — LLM-generated hierarchical document trees with entity enrichment (aircraft models, ATA codes, action keywords) and BM25Plus re-ranking for EASA AD documents. Applied entity-enriched tree nodes, iterative retrieval with sufficiency checks, query decomposition, normalised metadata filtering, and context budget management.

LangGraph · FAISS · BM25 · Azure OpenAI GPT-4o · FastAPI · Streamlit · asyncio

5-agent LangGraph pipeline (Planner → Retrieval → Writer → Validation → Assembly) with conditional routing and self-correcting loops for report sections scoring below 7/10 on tone, purpose, and quality. Concurrent async section processing with 70% FAISS dense + 30% BM25 sparse hybrid retrieval.

AWS SageMaker · S3 · Scikit-learn · Streamlit · Plotly · boto3 · FPDF · Kaleido

ML-powered Streamlit dashboard to predict HMW aggregate clearance across 5 purification steps using SageMaker-hosted models (Random Forest, Voting Ensemble). AWS S3 for centralised experimental data access and real-time SageMaker endpoint inference; embedded correlation heatmaps, clearance analytics, and scientific PDF export.