Product · 2026
Incident Desk
Should this alert wake an engineer up? Incident Desk asks TypeSafe's Jev, a decision model that returns typed answers with probabilities instead of text, four questions about an incident. Code rules then choose the next step: queue it for the owning team, ask a person to review it, or page on-call. Every incident is synthetic and routing is simulated.
The live product is private by design.
Every evaluation calls a paid model, so the running app is not publicly hosted. The demo above is the real product replaying real, recorded Jev answers, labeled as such on screen.
What it does
An incident report goes in: title, description, customer reports, recent changes, log lines, and a few monitor readings. Jev answers which team owns it, how much customer impact the evidence shows, whether waiting could make it worse, and whether the evidence is enough to decide. The app shows the proposed next step, the full probability distributions, and a trace of exactly which rule made the call.
How it's built
- Typed decisions, not prose. One request sends the incident and four independent Choice, Score, and Noul questions; Jev returns answers with probabilities, reached directly or through OpenRouter's System One API.
- Code owns the facts. Thresholds and durations are computed in code, and a hard rule pages on-call from confirmed monitor data regardless of what the model says, or whether it answers at all.
- Uncertainty goes to a person. Weak evidence, an unclear owner, or grey-zone urgency routes to human review. Malformed answers are rejected, never repaired.
- Measured, not assumed. An evaluation harness of 300 synthetic incidents in 60 families, split by family, runs the same code path as the app, with a rules-only baseline and spending caps on every paid run.