I build AI systems that replace manual enterprise workflows — from document intelligence to conversational analytics.
2 enterprise AI systems in production · 7-agent orchestration framework · 6-stage document extraction pipeline · 400+ research citations
Banks process thousands of Letters of Credit manually — reading PDFs page by page, copying fields into spreadsheets, cross-checking amendments by hand. I built a 6-stage AI extraction pipeline with a 4-tier routing strategy that reads, extracts, and validates structured data from LC documents in seconds. Human reviewers now verify AI output instead of doing the work themselves.
Business teams waited days for analysts to pull numbers from financial databases. I built a multi-agent conversational system — 7 specialist agents orchestrated by LangGraph — that translates plain English questions into live analytics with charts and tables. Users ask "Show me EBITDA trend for Q3" and get answers in seconds, not days.
AI agents that reason well can still act poorly. I designed a research framework that measures how reliably agents translate evidence into correct actions — quantifying the gap between "understanding" and "doing" that most evaluations miss.
Model upgrades, prompt changes, and retrieval tweaks silently degrade quality. I built a reproducible evaluation platform that detects regressions across model versions, prompt variants, and retrieval strategies — catching quality drift before it reaches users.
I take messy, manual enterprise processes and turn them into deployed AI systems. I own the full cycle — from understanding the business problem, to designing the agent architecture, to shipping production code, to sitting with users and watching what breaks. I don't hand off prototypes. I deploy and iterate until it works.
Python · Docker · FastAPI · SQL · LangGraph · Azure OpenAI · DuckDB · Redis · Azure
Currently: Researching evidence-to-action reliability in AI agents — measuring whether agents that reason correctly also act correctly.
Bangalore, India