The challenge
Rivermint had a card-issuing thesis, no engineers, and an eight-week fundraise window. They needed software that could process a real transaction, on real cards, by demo day, and stand up to investor scrutiny. A typical agency timeline of three to four months would have missed the round entirely.
Three constraints shaped the engagement:
- Financial correctness. Anything touching the ledger had to be verifiable, auditable, and free from the kind of off-by-one rounding bugs that get fintechs in trouble.
- PCI scope. We had to keep card data out of Rivermint's environment entirely, Stripe Issuing was the right call, but the integration had to be clean.
- Demo-day proof. Investors weren't going to read a Figma. They needed to swipe a card and see it work.
How we approached it
A 3-engineer pod paired with our QA lead, with a Hashorn PM coordinating across founder, engineering, and external (Stripe, KYC vendor). Week zero: a one-page product brief and a risk register. We modeled the ledger first, financial correctness above all, picked Stripe Issuing for the rails, and committed to a weekly Friday demo cadence.
AI handled scaffolding, type generation, and PR review. Humans owned the architecture, the financial-correctness tests, and every line that touched money.
What we shipped
Week 1–2, Foundation
- Multi-tenant SaaS skeleton with auth, RBAC, and audit logs
- Postgres ledger primitives with double-entry constraints enforced at the database level
- Stripe Issuing integration with idempotent webhook processing
Week 3–4, Customer surface
- Cardholder dashboard (transactions, balance, card management)
- Real-time transaction feed via Stripe webhooks → Postgres → SSE to the client
- Card creation flow with KYC vendor integration
Week 5, Operations
- Admin console: card freeze/unfreeze, manual transaction review, refund initiation
- 12-page operations runbook covering on-call, incident response, and Stripe escalation paths
- Playwright regression suite covering 14 critical flows; runs on every PR
Week 6, Hardening + launch
- Production deploy to AWS (ECS + RDS); blue/green rollout for the API
- Monitoring (Datadog) with alerts on ledger imbalance, webhook failure, and authentication anomalies
- Internal security review + handover
Outcomes
- Six weeks from kickoff to live card-issuing product. Demo day featured a live transaction processed end-to-end on stage.
- Zero production incidents in the first 30 days post-launch.
- Rivermint closed their seed round on time with the working demo as the centerpiece.
- Financial-correctness tests caught two real bugs that AI-generated code introduced, one off-by-one in fee calculation, one race condition on concurrent webhook processing. AI-augmented didn't mean AI-trusted.
What we'd repeat
The financial-correctness suite was the single highest-leverage decision of the engagement. Writing it before any business logic meant every change had to prove correctness, not just compile. AI is fast at generating code that compiles; only humans can write the assertions that actually catch financial bugs.