I'm an AI Product Manager who builds things. Not just roadmaps and PRDs — actual working products, with real AI wired in, shipped and running.

4+ years in enterprise tech — currently at American Express, previously MathCo. IIM Ranchi MBA on top of a B.Tech in Computer Science: I can walk into an engineering architecture review and then write the business case for the same decision. CommGuard exists to prove that gap is real.

My edge is that I sit at an uncommon intersection: I understand what AI can do technically, I understand what users actually need, and I can hold both truths in tension long enough to ship something that works in the real world — not just in a demo.

I've worked across the full AI product surface: LLM integration, evaluation frameworks, agentic workflows, compliance AI, enterprise workflow automation, data products, and statistical experimentation. The domain changes — the product thinking doesn't.

What I care about most:
  ├── When should the AI decide, and when should a human?
  ├── How do you evaluate AI outputs when "good" is subjective?
  ├── What does product-market fit look like when the model is probabilistic?
  └── How do you build AI products that scale without breaking trust?

What's inside the AI engine:

1. Capability is not a product. The model can do a lot of things. The PM's job is to decide which things are actually worth doing, for whom, and at what cost when they go wrong.

2. Evaluation is the hardest part. For traditional software, "does it work" is binary. For AI, it's a distribution. Defining what "good enough" means — and building the infra to measure it — is where most AI products succeed or fail.

3. The fallback is part of the product. Every AI feature has a failure mode. I design fallbacks first, not last. A keyword-rule engine isn't a consolation prize — it's the thing that keeps the product running when the model is unavailable, slow, or wrong.

4. Human-in-the-loop is a design decision, not a disclaimer. Where you put a human checkpoint — and what information you give them — determines whether AI actually helps or just creates a new bottleneck. I spend more time on this than on the AI itself.

5. Trust compounds slowly and breaks instantly. Users forgive slow. They don't forgive confidently wrong. AI products need to know when to hedge, when to explain uncertainty, and when to just not answer.

current:
  - How agentic AI changes the PM role — when the product is a loop, not a flow
  - Evaluation frameworks for LLM outputs that actually correlate with user value
  - RAG vs fine-tuning decision trees for enterprise knowledge products
  - Where AI governance becomes a moat, not just a cost center

exploring:
  - Multi-agent orchestration and what product design looks like when agents talk to agents
  - Structured output validation for AI in high-stakes decisions
  - Voice + LLM product surface — the interaction model is entirely different
  - AI-native B2C vs AI-embedded enterprise — very different product constraints

believe:
  - The best AI PMs can read a model card AND write a press release
  - A working prototype changes the conversation faster than a 50-slide deck
  - Responsible AI and fast AI are not opposites — safety is a feature, not a gate
  - The PM who understands the training data has an unfair advantage

Open to AI PM roles across FinTech · Enterprise SaaS · Developer Tools · Agentic AI · Regulated Industries