Founder project · built with a 20-agent team
Sep 2026 — Present
InvAI
An AI operations platform for DTF t-shirt print shops, built by an AI agent team I designed.
- repos
- 8
- repos
- typed API procedures
- 258
- typed API procedures
- tests
- ~1.8k
- tests
- eval cases
- 133
- eval cases
+Case study: problem, build, and caveats
Problem
Multi-marketplace print shops juggle orders from Etsy, Amazon, Shopify and TikTok Shop, plus gang sheets, labels and margins, across a dozen tools. I wanted to find out how far one engineer can go by directing a disciplined team of AI agents.
What I built
- Designed the agent operating system: 20 roles, 72 playbooks, path-ownership enforced by hooks, independent review, and an integration gate before every push
- Contract-first architecture: one typed oRPC/Zod package shared by the API, the dashboard and an offline-first tablet app across 8 repos
- Postgres row-level-security multi-tenancy (88 policies), a transactional outbox, and crash-safe exactly-once side effects for labels and purchase orders
- A Claude business-analyst agent with read-only tools, a 10-round loop cap, spend checks, a tool-output-is-data boundary, and an eval harness
Honest note: Agents wrote most of the code under my architecture, gates and review. It isn't deployed yet and has no pilot customers yet.
- TypeScript
- Node 24
- Hono
- oRPC
- PostgreSQL 17
- BullMQ
- React 19
- Python
- libvips
- Claude API
- AWS (SST)
