Most PMs talk to customers. I spent years being accountable to them. That's the lens I build everything through.
I build AI products for go-to-market teams and care most about whether they actually work. Most recently I was Product Lead at Petavue, building a multi-agent AI data analyst from nothing to roughly $400K in Year One revenue across 20 customers. The hard part was never getting an agent to produce an answer. It was knowing which answers to trust, which is why most of what I publish here is about evaluation, context design, and the unglamorous scaffolding that makes agents reliable enough to sell.
Right now I'm taking some time to figure out what to build next. There's a lot of noise around AI, and the easy move is to jump at the first interesting-sounding thing. I'd rather think it through.
- Agent evaluation: LLM-as-judge harnesses, rubric design, catching regressions when a prompt changes
- Context architecture: typed context layers, retrieval design, what an agent should and should not be told
- 0 to 1 GTM tooling: data products for go-to-market and customer success teams
- Building with AI: most of what is here was written with Claude Code, which is the point
| Project | What it is |
|---|---|
| write-like-me | A method for extracting someone's writing voice into a reusable agent skill. Captures by diffing rather than describing, and ships with an eval file it grades its own output against. |
Every step compounded into the next.
- MagicShapes: co-founded a 3D printing startup, designing and printing for students, engineering faculties, and R&D centers. Only ever 10 to 15 clients, but enough to learn what it takes to build something from scratch and sell it.
- Freshworks: customer success, a $2M+ ARR portfolio of 2,000+ accounts, churn cut from 6.5% to 3%. This is where I learned to work within what a product already does instead of reaching for something new. More useful than it sounds.
- Insent to ZoomInfo: joined Insent as the second person on the CS team, scaled it 10 to 50 customers, cut onboarding from 45 days to 14. After the ZoomInfo acquisition I moved into product and launched ZoomInfo Script, which cut implementation time 10x for 30,000+ customers. Not from a feature request, from knowing exactly where customers got stuck. Also grew FormComplete from $5M to $7M+ ARR and worked on WebSights match rates.
More about my work at shyamganapathy.com: writing on product work, agents, and the parts of building that do not make it into case studies.
When I'm not building, I'm usually training for the next marathon or outdoors somewhere. Lately picking up photography, badly but enthusiastically.