Staff Software Engineer · IntuitGenerative & Agentic AI

A point of view

The hard part of AI isn't the model. It's everything around it.

I architect the systems that make generative & agentic AI trustworthy in production — the evaluation harnesses, golden datasets, and guardrails that tell you when the model is wrong before your customers do. Fourteen years, five companies, tens of millions of users.

25M+

Monthly users served

4M+

AI datapoints evaluated / yr

14+

Years shipping at scale

1

Granted patent

Shipped at
Intuit/Microsoft/Docusign/Oracle/SAP

How I think about building AI

01 — 04

01

Models are commodities. Judgment is not.

The advantage was never the LLM you picked. It's the evaluation harness, the golden datasets, and the guardrails that tell you — reliably, daily — when the system is wrong.

02

Ground it in production reality.

RAG and agents earn trust only when they're measured against real customer utterances — not tidy benchmarks. The eval set is the product spec.

03

Make the black box explainable.

Per-utterance MRR, mAP and Hit-Rate, surfaced so a non-ML stakeholder can debug a bad score themselves. Trust scales when understanding does.

04

Automate the boring parts.

A good agentic workflow closes the vulnerability ticket and opens the PR for review — freeing humans for the ambiguous, high-judgment work only they can do.

Let's build

Hiring for AI, platform, or cloud at scale?

Whether you're building intelligent products, modernizing legacy systems, or evaluating me for a consulting engagement — the fastest way to find out is a conversation.

Start a conversation →