Data Science and GenAI Intern — RAG / LLM Applications at Munich RE

Built an internal GenAI Retrieval-Augmented Generation application over 1,000+ insurance claim files to enable claim assessors and Chief Medical Officers to query claimant documents, reducing assessment time by approximately 50%. Developed a Python-based hybrid retrieval pipeline using LangChain, ChromaDB, SQL, and internal LLM APIs to ingest, chunk, embed, retrieve, and answer questions over unstructured documents. Joined SQL databases with 1.2M+ claim records and 120+ columns to inject claimant metadata into vector stores and improve retrieval relevance. Implemented chunk deduplication using Levenshtein distance with >90% similarity threshold, reducing vector store size by ~15% and increasing retrieval diversity. Built automated indexing logic to detect newly uploaded claimant files and a metadata tracking system for automatic indexing. Proposed a holistic RAG evaluation framework using recall@K and precision@K, plus LLM-based faithfulness, relevance and answer reliability metrics.

AI/ML Engineer Intern: Retrieval-Augmented Generation (RAG) at BackPro AI

Developed document ingestion and processing pipelines (PDF/DOCX parsing, chunking, pgvector storage) in a Dockerised Python monorepo. Used frameworks such as fitz, PDFPlumber and docx2python to extract structured data (tables and images) from PDFs and Word documents for processing into RAG pipelines.