The same photo with the Berry Evening look applied by the try-on

BeforeAfter

AI you cantry on.

Job search

Open to my next role.

Senior AI product roles, especially agents and conversational products. Based in Bengaluru, open to remote. I own products end to end, from direction and architecture to evaluation and go-to-market.

Senior AI Product Manager
AI agents
Conversational AI
0-to-1
Evals & guardrails
Remote friendly

Email me Resume (PDF) LinkedIn

4.7s → 1.2sMedian time to first token on Convexa, my AI agent at Nagent

500+Production conversations handled, under 10% left unanswered

~85%Visibility loss on deeper skin tones I found in a naive render, then fixed

9 yrsIn product, from consumer commerce to gaming to AI agents

Selected work

Things I built, with the receipts.

Two projects, each with its own page. Numbers are marked with how they were measured.

Live demo · Prototype

Live AI Makeup Try-On

Try lip, eye and cheek shades on your own face, live from your webcam, then see the look in daylight, candlelight or a club. Built as an AI product case study.

On-device. Face tracking runs in your browser. Video and photos are never uploaded.
Fairness, measured. A darkening-only blend lost about 85% of its visibility from the lightest to the darkest skin tone (ΔE 60.9 → 11.0). The shipped 3-pass render holds a 24 to 32 ΔE band across all 10 Monk Skin Tone references.
Rules on top of ML, on purpose. Shade matching is a transparent heuristic on a face-tracking model, not a trained model, and the case study says why.

Honest status: a prototype with a demo catalog. The fairness result is a synthetic colour test of the rendering math, not real faces. The real-user evaluation is still to do.

Try the live demo Read the case study

Product · Nagent AI

Convexa, a conversational execution agent

Not a "RAG chatbot". An agent that qualifies a lead, syncs the CRM and hands off scheduling in one conversation. I own it end to end: direction, architecture, evaluation, pricing and go-to-market.

Faster and cheaper together. Median time to first token cut from 4.7s to 1.2s by moving to hybrid BM25 + semantic retrieval, bounding context to five turns and caching the compiled system prompt.
Strict grounding. The agent says "I don't know" instead of guessing, with a three-level guardrail precedence and a golden-dataset eval suite (regression + LLM-as-judge) gating every release.
In production. First paying enterprise customer, 500+ conversations handled, under 10% unanswered.

See the Convexa project Full teardown

How I work

Judgment first, then the model.

The AI part is rarely the hard part. Deciding where to use it, how to prove it works, and what to do when it's wrong is.

Pick problems AI is actually needed for

I start from the user's blocker, then ask if something simpler works. In the try-on, shade matching is rules on a tracker, not a trained model, and I can explain what evidence would change that.

Evaluate before you ship

Golden datasets and regression checks gate releases on Convexa. On the try-on, a colour-parity test caught a fairness bug before any user saw it.

Name the guardrails

"Safety filters" isn't a control. Lighting checks, explicit "rough guess" labels, strict grounding and on-device processing are.

Say what isn't done

Every project here states what was measured and what wasn't. A number with no method behind it doesn't belong on a portfolio.

About

Product, now with AI at the core.

I'm a senior product manager with nine years across commerce, gaming and AI. I scaled Call Break into a top-3 revenue game at Mobile Premier League, leading a 15-person team, and I'm now building AI agents at Nagent.

I build to learn. The projects above are working software I designed, tested and can walk you through, including the parts that broke.

Based in Bengaluru, open to remote. Looking for senior AI product roles, especially agents and conversational products.