The problem
Everything at work is instant except getting paid back. You spend on behalf of your team, then you scan receipts, fill in forms, chase approvals, and wait weeks for a bank transfer.
What I built
Reimburse AI, my first company, incorporated in the US as Reimburse Inc. I was co-founder and CTO, and owned the product and the technical stack.
- Document AI. A GPT-4o Vision pipeline that reads receipts (amounts, dates, vendors) and checks each expense against company policy, with fraud detection built in.
- Evals. Test sets for ambiguous, low-quality, and adversarial receipts, plus a human-review step for low-confidence cases.
- Instant, non-custodial payouts. Approved expenses were paid in USDC straight from the company’s own Safe wallet through a spending-limited AllowanceModule, so we never held company funds. It ran on Base, Avalanche, Polygon, and Ethereum, and it was gasless for employees.
- The stack. Python, FastAPI, OpenAI, TypeScript, Next.js, and Postgres.
Traction
- 9 signed letters of intent before launch: ntmax, 33audits, College DAO, Drip Pay, Oregon Blockchain, the Lebanon Blockchain community, UW Blockchain, UCI Blockchain, and Chicago Blockchain.
- Top 10% of Y Combinator applications.
Market insight
In India, UPI’s zero-MDR economics break the card-interchange model that funds companies like Ramp. The defensible layer is AI verification and auto-approval, not the card.
Why we stopped
The product helped employees, but companies were the buyers. In the US, companies were already on Ramp and corporate cards. We talked to many customers, investors and other teams, and it was clear they would not move their whole reimbursement cycle to us. So we shut it down.
The lesson: the user wasn’t the buyer. Build for the person who pays, not only the person who uses it.
What came next
I co-founded SnowMind.