Offramp, built in five days (19–23 Oct 2026)
In plain words: Offramp moves an app's AI calls to cheaper NVIDIA Nemotron models, only where tests prove they still work, and opens a pull request with the proof.
How it evolved
- 19 Oct: the whole pipeline, end to end. Scanner (Python AST) to find every OpenAI and Anthropic call, eval runner with code scorers and a Nemotron 3 Ultra judge, an Iterator that rewrites prompts when a cheaper tier nearly passes, per-site verdicts, and a patcher that runs the repo's own tests in a network-off Docker sandbox. Every model call goes to Nebius Token Factory.
- 20 Oct: models propose, code decides. Nemotron 3 Super's rewrites of app code were rejected on 5 of 6 sites, so app-code changes moved into deterministic code and the model now edits only tests and fakes. The demo went from 1 patched site to 3, and later to 5.
- 21 Oct: evidence judges can check. A report and replay anyone can open without keys, a keyless scan at /scan, a scan of 30 public repos, and prices checked against their source pages with Tavily.
- 22 Oct: a real repo we didn't write. On a fork of mnemox-ai/idea-reality-mcp, Offramp found five bugs in itself (all fixed, with tests), then moved 2 of 5 calls to Super, kept 2 on Claude and sent 1 to a person. Final demo run: 6 calls checked, 5 migrated in one PR, est. −95% cost on those 5, the repo's CI green.
- 23 Oct: final review and polish. Replays that stay exact after a retry, PR text that says exactly which tests ran, the demo video, and this submission.
Try it: https://offramp-seven.vercel.app (no login, no keys) · Video: https://youtu.be/dECD_2mqnME · Code: https://github.com/vj-io/offramp
Total Token Factory spend: $4.13 of our $25 credit. Feedback welcome in the comments!
Log in or sign up for Devpost to join the conversation.