SpecForge

Inspiration

Writing a good app starts with a good spec — but specs rot, and building from them takes a team. We wanted the loop itself to be the product: describe what you want in plain English, and get back a repo that passes its own tests.

What it does

SpecForge takes a plain-English specification and produces a working, test-verified application repo: researched, generated, refined, and hostile-QA'd until the tests pass. A refinement loop iterates against real execution, and the hostile-QA regression guards the output — 19/19 green — before the repo ships as a ZIP. Every generated project also ships a component inventory: each component, its version, and its dependencies, tracked from the real project files, plus a COMPONENTS.md in every workspace.

How we built it

• Models: NVIDIA Nemotron via Nebius Token Factory — fast pass on nvidia/nemotron-3-super-120b-a12b, reasoning on nvidia/Nemotron-3-Ultra-550b-a55b (live catalog IDs, verified end-to-end).
• Grounding: Tavily research so generated apps reflect real-world facts, not just the model's priors.
• Loop: generate, run tests, refine, with a hostile-QA regression suite (19/19 green, real execution). Two real bugs caught and fixed this way: a 404 root route when the frontend wasn't built, and __pycache__ pollution in ZIP exports.
• Studio: a 21st.dev/shadcn-style studio UI with a live component tracker, backed by a FastAPI backend (42/42 backend tests green).
• Verified end-to-end: GO FOR LIVE DEMO at 23/23 green — a full spec-to-repo run in 35.59s, under $0.03 in inference cost.

Challenges we ran into

Two real ones the QA loop caught: the app 404'd its root route when no frontend was built, and generated ZIPs shipped __pycache__ clutter. Both are fixed and covered by the regression suite. Honest gap: without a Nebius Token Factory key the app falls back to mock mode — it can't demonstrate real Nebius-infra plus NVIDIA-model usage, so participant credits and the key matter.

Accomplishments that we're proud of

• 23/23 tests green in real mode; 19/19 hostile-QA regression green.
• A full spec-to-repo pipeline in about 36 seconds for under $0.03.
• The refinement loop that fixes its own bugs instead of hiding them, plus a component inventory in every generated project.

What we learned

The test loop is the product. A generator that only writes code will happily ship broken code; a generator that must pass its own tests, and gets re-run until it does, is the difference between a demo and a deliverable.

Built With

• fastapi
• nebius-token-factory
• nvidia-nemotron
• python
• react
• tavily
• typescript
• vite