What I Build

I was one of the earliest people on the team, and I run engineering end to end here: strategy, architecture, budget, hiring, delivery, all of it. We've scaled to about 1.3 million monthly users and shipped 2 pivots, the second a full rebuild in 6 weeks with 3 engineers.

Keeping the economics honest matters as much as the engineering. Tech is about 18% of what we spend acquiring a customer, and I've kept the infra deliberately boring: containers on ECS Fargate, GitHub Actions for CI/CD, so the team spends its time building instead of fighting the stack. Deploys run 2-3 times a day per engineer, 8-10 minutes from merge to production behind feature flags, with 5-minute rollbacks when something's off.

Underneath it all is an agent platform we built: it keeps 7.9K product lists freshly ranked (217K products across India and the US) at under a dollar a list, with LLM judges catching most mistakes before they publish. That same platform, the product catalog, pricing, context graph, and commerce graph it maintains, is what now powers selling over WhatsApp, which we're piloting as a key channel.

I was also part of a team of 2 that started Buffy, a B2B sales agent for Shopify stores. It worked, just not as a business, so we shut it down.

None of this needed a bigger team. I've held the budget flat for 4 years running, betting on density over headcount, and replaced most of the ceremony with written specs and design-first reviews. Idea to proof of concept takes a day or 2 now, not weeks; new hires are shipping to production inside their first week.

I stay hands-on across infra, agents, and UI here. The second pivot, rebuilding flash.co as a marketplace where agents do what a storefront used to, took 3 of us 6 weeks. I built the WhatsApp agent we're now piloting as a growing channel for bringing people in, and rebuilt our email infrastructure: 50,000 addresses to about 1.5 million, cost per million emails down more than 70%.

The part I'm proudest of is the agent platform underneath it all: a pipeline that keeps around 7.9K product lists ranked and current (nearly 217K products), pulling from 5 research sources, validating with LLMs, running nightly supervisors to catch what slipped. Only about 1 in 4 flagged removals turns out to be right, which tells you the validators are doing real work, not rubber-stamping. Judges correct most builds before they go live, and automated fixes land right the vast majority of the time.

Under the hood, I moved the agent stack from an in-house harness to LangGraph and then to Mastra, after trying CrewAI and a few others, and built real durability on top of it: idempotent steps, heartbeat-based ownership, orphan recovery. We've pulled back over 300 runs across redeploys that would otherwise just be lost. I re-benchmark the model router every few months and pin tasks to whatever's cheapest and good enough; plenty of this doesn't need an LLM at all.

I've also kept the team small on purpose. Budget's been flat for 4 years, because I'd rather have a few people who are excellent than many who are fine. Specs over meetings, design before code: idea to working prototype takes a day or 2 now.