I build things that make AI agents trustworthy, and I teach what I learn.
Former backend and data engineer at Adobe and Yahoo. Today I lead product for a marketing data and AI platform, where I took a production MCP server to general availability. MBA, Cornell. The work is below, sorted into three rooms.
Room 1 · the front door
My expertise
Open-source tools for trusting agents with real business data. Each ships with an eval and a list of what it doesn't prove.
Is the number right?measurement-md ↗Define a metric once and serve it to agents over MCP, a CLI, a query gate and a Snowflake semantic view.62 of 62 planted mistakes refuseddata-layer-interop-checker ↗Finds broken joins in a marketing data stack.100% detection, 0 false alarms
Is the agent allowed to do this?warrant ↗Scoped, revocable authority for agents acting on a person's behalf.16 of 16 attacks denied for the right reasonagent-ledger ↗An audit trail of what an agent did, and on whose authority.
Does it hold when many agents run at once?quorum ↗A chaos harness for agent fleets, and the fixes that survive it.384 lost updates → 0
Is the platform ready, and did the change help?agent-readiness-scanner ↗Scores a platform on whether agents can use it.prompt-bench-cli ↗A golden-set harness for prompt and model changes, with a regression gate.
Room 2
Viao↗Done-for-you AI for local service businesses, mapped across 20 verticals.First real test: 28 roofing companies, no sales yetPorino↗A technical reference on agentic systems, stream processing and ML infrastructure.Things I've built and tested
Real products and real tests, including the ones that haven't worked yet.
BhashaCustomer messages in each customer's own Indian language, for lenders working to a trilingual-communication rule.Specified, not built
LifeBookA private place to remember everyone from each chapter of your life, before any personal CRM can help.Built for myself; kept private on purpose
Idea inbox→Ideas I have thought through, with the riskiest assumption written down, including the ones I decided not to build.Room 3
Life lessons→One interactive page per life lesson, written for my son and kids around 10 to 14. First lesson: Gardeners & Miners.AI Fellows↗A free six-week fellowship: practitioners review students' real work every week.Threshold↗Mentorship for students in small-town India from diaspora professionals.Paar↗Bengali literature in English translation, starting with books that exist nowhere online.AI Systems Thinking for PMs↗A free lightning lesson on Maven.Teaching and giving back
For students, kids and anyone getting started.
I write down what I learn.
Essay archive145essays on product craft, platform strategy and scaling ↗
Peer-reviewed · co-author · IEEE ADDA 2025A Heuristic-First Evaluation Framework for Marketing AI Agents: Judging Quality, Actionability, and TrustHow to judge whether a marketing agent's answer is good enough to act on.
Peer-reviewed · co-author · IEEE SoutheastCon 2026Explainable AI for Marketing Intelligence: Interpretable Multi-Agent Systems for Data-Driven Decision SupportMaking multi-agent recommendations explainable to the people who act on them.
I studied 46 industries. The same five problems kept showing up.
Vertical AI is not a chatbot per industry. It is software that joins the money spent to the outcome that matters, inside each industry's own system of record.
AThe join problemThe outcome lives in a system the ad tools can't see.
BCapacity caps revenueMore leads to a shop that can't staff them is waste.
CThe calendar confoundsPromo weeks, tax season and term dates distort every benchmark.
DThe wrong conversionSign-ups get counted; customers who stay pay the bills.
ESomeone owns the middleMarketplaces and aggregators rent you your own customers.