I deploy AI where it has to work.

First-of-its-kind production agentic AI workflow for a Fortune 500 enterprise. Embedded with the document review team, identified that reviewers were burning significant hours per week on repetitive policy checks. Designed and shipped an event-driven pipeline: API Gateway → Step Functions → parallel Lambda tools (PDF parsing, chunk planning, classification, policy validation) → RAG grounding on AWS Kendra GenAI Index. Every output traceable via correlation IDs for compliance audits.

Ed-tech platform for Maharashtra MSCE scholarship exam prep, built from a gap discovered by embedding with parents and exam coordinators in the scholarship ecosystem. Parent-as-gateway model with child profiles, tiered access, and 124-test end-to-end testing plan. Generically extensible exam-category configuration enables rapid expansion to new exam verticals. Supabase MCP integration with Claude Code for AI-assisted delivery.

Spiritual mantra-tracking PWA built for practitioners of Vedic disciplines. Originally architected on AWS (Lambda, RDS Proxy, Cognito JWT, SAM) before owning the platform-migration decision to Firebase. Implemented global admin via Firebase Custom Claims and Razorpay subscription/donation flow with tiered pricing.

Open-source Python pipeline I donated to a charitable healthcare event that needed to digitize 15,000+ handwritten prescription PDFs in 48 hours. Built with Claude Sonnet (Anthropic API), Google Drive API, ThreadPoolExecutor concurrency, and a Tkinter operator GUI so non-engineer volunteers could run it on any Windows laptop. Released publicly so other charities and clinics can reuse the pipeline.

Open-source Python package and CLI for automating handwritten document workflows. Converts forms, notes, and scanned paperwork into structured JSON using vision LLMs and optional schema guidance.

An MCP server that gives any AI agent zero-config time-series forecasting. Plugs Google's TimesFM 2.5 foundation model — plus a pure-NumPy statistical baseline — into Claude, Cursor, or any MCP client via one install line. The agent calls a forecast tool, gets point predictions, confidence bands, and a trend/seasonality summary, then writes the recommendation itself. Includes a backtest tool reporting MAE/sMAPE on held-out data so forecasts are validated before they're trusted.