Member since July 2026
I'm a Senior Product Manager at Microsoft with over 10 years of experience building products people rely on. I work on grounding and retrieval, the infrastructure that decides whether an AI's answer is trustworthy enough to act on, and I've contributed to Copilot grounding and Agent 365, with launches featured at Microsoft Ignite keynotes. Before Microsoft, I built products at Amazon and American Express, growing from data and analytics into AI product leadership. I speak and write about what I think is the defining skill of this era: not how AI works, but how we decide when to trust it. I write about AI trust and evaluation for O'Reilly Radar and CIO.com, and I publish technical papers on AI systems, including with IEEE. My talks cover AI product management, building trust in AI systems, decision-making when machines give you the answer, the evolving PM role, and leading without authority. I bring practitioner depth rather than trend commentary, and I favor hands-on sessions where the audience works through real problems instead of watching slides.
The moment someone realizes they've been trusting an answer they never actually checked.
I love watching that land in a room. We've all done it: the AI gives you something clean and confident, and you quietly stop asking questions. Nobody talks about it because it feels embarrassing, but it's happening in every company right now, at every level.
What lights me up is turning that discomfort into a skill. Judgment isn't a personality trait you either have or don't. It's learnable, and it's the most valuable thing a product person can build right now. Getting to hand someone a framework they can use the next morning and watching them realize they're allowed to push back on the machine, is the best part of my job.
The most instructive AI failures are the ones where nothing looks broken.
I work on the systems that decide whether an AI's answer is trustworthy enough to act on, and the pattern I keep coming back to is this: every control passes, every log looks clean, every metric is green. The answer is confident, well-worded, and wrong. Nobody catches it because there was never a definition of "right" to catch it against.
I like telling this from the stage because the room always gets very quiet. Everyone recognizes it. They've all shipped something they couldn't actually verify, and most have never said that out loud.
The story I build from there is about what changed when agents started acting rather than answering. A wrong output used to be an annoyance. Now it's a sent email, a booked meeting, a modified record. The blast radius changed and our habits didn't. That's the conversation I want to have with a room.
I grew up on the Andaman Islands, a remote archipelago in the Bay of Bengal, a long way from the rooms where technology gets built. I came to the US for a master's in Computer Science and started my career in data.
For years my job was turning messy data into something a business could actually decide with. Different companies, different industries, same question underneath: how do you know this is true, and what do you do if you're not sure?
When I moved into product, and then into AI product at Microsoft, I expected that question to get easier. It got harder. We built systems that could answer anything, and somewhere in there we stopped asking whether the answer was right. The confidence of the output started standing in for the correctness of it.
That's when it clicked for me. The hard problem was never capability. It's judgment. And the people who most need that skill aren't researchers, they're the product managers, leads, and engineers who have to make the call and live with it.
I think growing up far from the center of things gave me a useful habit: I never assumed the room knew better than the evidence. That instinct turned out to be the whole job.
So that's the work now. Building the infrastructure that makes AI trustworthy and helping the people who use it learn when to lean in and when to push back. Same question I started with, much higher stakes.
The most instructive AI failures are the ones where nothing looks broken.
I work on the systems that decide whether an AI's answer is trustworthy enough to act on, and the pattern I keep coming back to is this: every control passes, every log looks clean, every metric is green. The answer is confident, well-worded, and wrong. Nobody catches it because there was never a definition of "right" to catch it against.
I like telling this from the stage because the room always gets very quiet. Everyone recognizes it. They've all shipped something they couldn't actually verify, and most have never said that out loud.
The story I build from there is about what changed when agents started acting rather than answering. A wrong output used to be an annoyance. Now it's a sent email, a booked meeting, a modified record. The blast radius changed and our habits didn't. That's the conversation I want to have with a room.
I grew up on the Andaman Islands, a remote archipelago in the Bay of Bengal, a long way from the rooms where technology gets built. I came to the US for a master's in Computer Science and started my career in data.
For years my job was turning messy data into something a business could actually decide with. Different companies, different industries, same question underneath: how do you know this is true, and what do you do if you're not sure?
When I moved into product, and then into AI product at Microsoft, I expected that question to get easier. It got harder. We built systems that could answer anything, and somewhere in there we stopped asking whether the answer was right. The confidence of the output started standing in for the correctness of it.
That's when it clicked for me. The hard problem was never capability. It's judgment. And the people who most need that skill aren't researchers, they're the product managers, leads, and engineers who have to make the call and live with it.
I think growing up far from the center of things gave me a useful habit: I never assumed the room knew better than the evidence. That instinct turned out to be the whole job.
So that's the work now. Building the infrastructure that makes AI trustworthy and helping the people who use it learn when to lean in and when to push back. Same question I started with, much higher stakes.