What it does
Relay turns incident notes and logs into a handoff draft in which every factual claim carries an exact source excerpt. Claims with missing sources or fabricated excerpts block acceptance. A named reviewer must check every claim before the handoff can be accepted and exported as JSON. Suggested actions are never executed.
Why
AI-written incident handoffs read confidently even when details are invented. Relay makes provenance checkable (exact-quote validation) while being honest that a matching quote shows where text came from, not that it is true: a human reviews each claim.
How it uses Nebius and NVIDIA
Generation runs through Nebius Token Factory (OpenAI-compatible API) with NVIDIA Nemotron-3-Nano-30B-A3B, prompted for strict JSON claims with quotes; a local validator rejects any quote that is not a substring of the sources. Source text is treated as untrusted (a sample injected instruction is ignored and flagged).
Honest status
Local prototype using synthetic incidents. Real Nebius + Nemotron calls were made Oct 19 (three smoke requests plus one browser retry); outputs were inspected by an agent, not independently validated by a human. No measured time savings are claimed. The default demo runs offline with no key.
Run
python3 -m relay (offline) or python3 -m relay.web (browser review demo); --live with NEBIUS_API_KEY for Nemotron. 20 tests. Standard library only. MIT.
Log in or sign up for Devpost to join the conversation.