Selected work
Products & agents I've shipped.
How I think about building AI-native products — the problem, the agent and eval design, the tradeoffs I made, and what moved.
- 01 5x Product velocity lift 1 Slack msg → PRD, mockup & Linear ticket Minutes Insight → reviewable artifact (was hours)
Building a 5x PM agent on Claude Code
An always-on autonomous PM agent: one Slack message in, a PRD, a mockup, and a Linear ticket out — built on Claude Code + MCP with parallel research sub-agents and full observability. - 02 4x Per-user operational efficiency 40% Faster turnaround time
Automating fax-based referral intake with an AI agent
Specialty clinics still run referrals over fax. An AI intake agent cut turnaround 40% and 4x'd per-user operational efficiency. - 03 +5% Trial-to-paid conversion (2 mo) 5 in 7 Activation metric (notes / days) 3.5% → <1% Incident rate, cut in parallel
Rebuilding Supanote's PLG activation path
Signup→subscription sat flat at ~12% for months. I reframed the quarter around conversion, found the real leak — setup, not the demo note — and lifted trial-to-paid +5% in two months. - 04 2.2x ARR in 3 months 0 → 1 Launched the product line Eval-gated Regressions caught pre-production
Re-architecting benefits verification into an agentic system
Took a brittle rule-based enterprise workflow to a harness-based agent orchestration — 2.2x ARR in 3 months, with eval-gated reliability for an enterprise bar.