About

I'm Yash Gadodia, an applied AI product manager based in Singapore. I came from engineering: five years writing code before that, on platform systems at Ninja Van (notification infra, voice calling, parcel booking) and an NLP/RAG platform at Synthesis Partners for Fortune 500 clients. I moved into product because the most interesting decisions in AI products are not implementation decisions, they are decisions about what the agent should refuse to do, how you measure whether it is doing the right thing, and which job description it is actually filling.

Today I'm the founding PM at Voltade, where we ship two 0-to-1 platform products: Studio (a no-code agent builder) and Envoy (a conversational CRM). 100+ SME accounts, 230K+ AI interactions a day, 99.65% task success rate. I am closest to the user when I am on a customer's WhatsApp watching the agent answer Kai from the bakery.

#How I think about AI products

These three sit underneath every decision I make. They are not slogans, they are how I argue with myself when the work is hard.

Sell the work, not the tool. The best AI products deliver finished outcomes, not another interface to manage. When we built Envoy we deliberately did not build a "ticketing system." SMEs do not want another inbox to triage. They want a thing that already answered the customer and gave them an audit log. The structural choice (agent drafts, human approves, message goes out) came from this, not from a feature spec.

Redesign around the capability. Most AI features fail because they automate old workflows instead of asking what new workflows the technology makes possible. The WhatsApp-first decision at Voltade is the example I keep coming back to. The conventional wisdom is "meet the user where they are" which means a web dashboard with WhatsApp as a channel. We inverted it. The agent lives in WhatsApp because that's where the customer already is. The dashboard is the audit surface, not the product. That single inversion is the difference between "another CRM" and "the thing that already answered."

Evaluation is product work. If you cannot measure whether your agent is doing the right thing, you have not finished the product. I built WIMAUT after an OpenClaw cron job silently burned $300 because no one was watching. Since then, every agent we ship has a three-layer eval harness, a failure taxonomy, and a drift detector before it is allowed to talk to real customers. I wrote about how we do this in more detail.

#What I'm optimising for

I want to do this work at a place where evaluation is the product and not the audit. Frontier labs that are shipping applied AI into production at scale (Anthropic, OpenAI, Manus) are the closest fit, alongside Singapore-anchored teams (OGP) where I can stay in the region and still work on consequential AI products. I wrote down what I'd build first if I were starting an applied AI lab in this region.

The shape of the role matters more than the title. Founding PM, agent PM, model behaviour PM, deployed PM. Anywhere the question "is this agent actually doing its job" is the centre of the work.

#Now

What I'm running today, from personal agents (Claudia, Lawrence) to Voltade products (Studio, Envoy), is on Projects. Earlier: platform engineering at Ninja Van (100M+ events/day), co-founded AfterClass (50K+ users, still active), built a DTC consumer brand at The Bon Pet.

#Elsewhere on the site

  • Frameworks: the named patterns I keep using when shipping AI products. Eval, behaviour spec, scope-and-state, self-learning loop, cost tiering.
  • Stack: the hardware, software, and agent infrastructure I run daily.

#Contact

If you want to talk, reach me at pirsquare.yash@gmail.com. I read everything.