At QuantaSight, I architected a LangGraph multi-agent research assistant for pharma clients that returns cited, infographic-backed answers in about 20 seconds instead of hours of manual searching. I also developed clinical claim verification and presentation-generation pipelines that replaced days of analyst work.
I built a KOL intelligence chatbot that turns natural-language questions into queries over CRM records, then returns a ranked table, chart, and summary. To improve answer relevance, I combined BM25 search, cross-encoder reranking, and query rewriting; I also fine-tuned Gemma 3 and SmolLM2 with LoRA, with the best classification model reaching 89% F1.
Earlier, at Vaultedge, I fine-tuned a table-detection model that raised invoice accuracy from 72% to 95%, and developed a cheque-classification pipeline that reached 98% accuracy. At The Web Blinders, I worked on sentiment analysis, transaction-based inventory insights, and a recommendation engine.
