Sirv AI Studio is an AI product-content platform for e-commerce: 30+ AI tools, batch workflows, supplier intake, and safe publishing to Shopify. I designed and built it. Solo for the first twelve weeks, and still the author of roughly 73% of every commit since. Max Wish built the design system and the data grid; Veniamin Krachun built the QA machine. Below is the build record, straight from the git log.
Rendered from real per-day commit counts. Seven months, five days off. The hot streak from April onward is the platform hardening for production: billing, entitlements, supplier portal, publish safety.
| Dec 2 | First commit, December 2, 2025. Empty repo to production platform in seven months. | git log --reverse |
| 7,515 | Commits in the repo. 5,500 are mine, about 73%. Four engineers have joined since. | git shortlog -sn |
| 12 wks | Solo. Every commit in the first twelve weeks is mine or my AI agents’. The next engineer landed Feb 26. | git log --author |
| 254 | Database migrations, each one a schema change that shipped. | ls drizzle/ |
| 3,695 | Test files behind a blocking quality gate. Velocity without tests is just typing. | find . -name '*.test.*' -o -name '*.spec.*' |
| 39/day | My commit rate over the last 30 days. Month seven, faster than month one. | git log --since='30 days' |
| 30+ | AI tools live in production, from background removal to full workflow orchestration. | sirv.studio |
E-commerce teams get product images from suppliers, photographers, and old catalogs. The content is messy, the fixes are manual, and publishing a wrong image to a live store is expensive. Studio turns that into one controlled loop:
Under the loop: a product-linked asset library, supplier upload portals with validation and SFTP intake, readiness scoring, a visual workflow builder, and an API/MCP layer so AI agents can run the same operations the UI does, inside the same permission and approval rules.
A production MCP server (45 tools, published on npm) and OpenAPI layer where AI agents run real product-content workflows, and every operation respects credits, permissions, review gates, and publish safety. Agents get the same rules as humans, enforced in the same place.
Writes to live Shopify catalogs are guarded by source drift detection, idempotency keys, per-product status, preserved originals, and one-click rollback. Partial failure is a designed state, not a surprise.
Suppliers get scoped upload portals and SFTP drops. Files are validated on arrival against filenames, SKUs, slots, and specs, repaired by AI autofix, and held in an approval queue enforced at the database layer. Supplier content cannot reach a live store without a human yes.
A visual pipeline builder compiling to DAG execution over durable background jobs, with 36 step types, live SSE progress, human review gates, and credit metering per step.
The honest answer: I run a fleet of AI coding agents the way a lead runs a team. Test-first on every behavior change. A blocking quality gate on every stop. Adversarial review agents that try to break each change before it lands. Ratchets that stop files from growing back.
The agents type faster than I do. The architecture, the judgment, and the taste are mine, and the 3,695 test files are why I can trust the throughput. In 2026 every engineer has the same models. The difference is whether your system turns that speed into a product or into a mess.
The codebase is private, so you cannot run the commands in the ledger yourself. I will happily walk through the git history live in any interview, screen shared, no preparation. The product needs no permission at all: sirv.studio is public and free to try.