At Miimansa, I productionized a LightGBM classifier for deduplicating clinical records, reaching 97.35% precision with zero duplicate-collision or data-corruption errors. I also deployed Gemma 4 (26B) with FP8 and KV-cache quantization for a live LLM-ensemble entity-extraction service.
I built core stages of a 9-service NLP pipeline that turns hospital discharge summaries into structured OMOP CDM records and a longitudinal patient knowledge graph. Its extraction and standardization layer handles seven entity types and combines vector and lexical methods.
In my projects, I built a credit default risk model on data from 300K+ applicants and a support-ticket triage agent that improved escalation recall from 14% to 85%. I'm also a co-author of GuardPaint, under revise-and-resubmit at ACL ARR 2026, and the first author of NeuroCrypt.
