I build production AI products from scratch by directing AI coding agents, the same way I lead engineers, then prove they work. Eight years shipping healthcare agents and eval-first systems that cite a verified source or abstain, never a confident guess. Available to US teams as an advisor, a fractional lead, or a hands-on builder.
SUBJECTA citeproof receipt. Pick a claim and watch it find the exact supporting line, or abstain when no line supports it.
Interactive explainer. The real run is shown below.
Destructive fishing practices, including blast fishing and cyanide fishing, damage reef structures and reduce fish populations. Rising ocean temperatures cause coral bleaching, in which corals expel the symbiotic algae living in their tissues. Reefs also buffer coastlines from storm surge and erosion, protecting communities near the shore.
When a claim has no support in the snapshot it is dropped, not guessed. That refusal is the whole product: a local entailment gate against a verified source, anchored to a verbatim span. Abstain over guess.
Three ways in, one conversation. Tell me the problem on a call and we will pick the lightest engagement that solves it.
Retainer or sprint
Is your agentic system working, or just demoing well? I build the eval harness that measures it and review the architecture around it.
Part-time, monthly
Embed as your AI engineering lead. I have run teams and shipped production agentic AI, so yours skips the traps and ships sooner.
Fixed scope
You have a scoped problem now. I build it from scratch with agentic development, the RAG pipeline, the agent, or the full product, ship it, and hand it over with the evals that keep it honest.
For the right team, also open to a senior full-time or fractional leadership seat. US-remote, EST overlap, no visa required.
Book a 30-min callProduction systems I built and led, some end to end with agentic development. The code is proprietary, so the numbers are the proof.
Agentic Chrome extension + conversational RAG agent for healthcare credentialing
Serverless AI verification across all 50 US states
ML fingerprint synthesis + human-behavior simulation, Forage AI
Production-grade primitives, all Apache-2.0. Click any repo. Every link resolves.
Fetch, but it tells you the truth.
HTTP returns bytes. veriscrape returns a portable trust verdict alongside them, so downstream systems can decide whether to trust the page.
Research that won't cite what it can't verify.
A local deep-research agent. Each cited claim links to a receipt: the page snapshot, highlighted to the supporting line. Unsupported claims are dropped, not guessed.
Turn flat PDFs into real, fillable AcroForms.
Most PDF tools render forms or extract data, not both deterministically. acroforge builds true AcroForms with named fields and calculation logic, zero copyleft surface.
Sound judgment from every document.
An AI document workspace for US professionals. Upload anything, it pulls out the parties, dates, amounts, and obligations, then answers questions with citations to the exact document and page. Built solo, from scratch, with agentic development.
You are not just buying what I shipped. You are buying how I operate when the model is confidently wrong in production.
If there is no metric for the change, it does not ship. Golden sets land before code.
Working software in front of stakeholders every Friday. Surprises die in the demo, not in production.
Postgres, Docker, FastAPI, Lambda. The clever part stays in the model layer, not the plumbing.
Real EST overlap, plus progress while your team sleeps. Written decisions over meetings, so the record stays the source of truth.
Lead engineering for a production healthcare credentialing platform: Django backend, serverless verification across 50 states, and Assured Copilot, the agentic RAG product I built end to end.
Automation architecture and data engineering for enterprise clients.
Built anti-detection extraction at scale, 1M+ profiles/day, defeating Cloudflare and bot-detection.
Led 8 engineers building an NLP-driven real-time social listening platform.
Legal-court, USPTO patent, and automotive data pipelines. Where the scraping craft started.
B.Tech, Electronics & Communication, Bharati Vidyapeeth College of Engineering, New Delhi, 2013-2017
If you are building agentic AI and the cost of being confidently wrong is high, that is exactly my work. I reply within 24 hours.
Based in New Delhi, US-remote, EST overlap, available for contract and consulting, no visa required.