What it is
A support triage service that treats Jev's probabilities as inputs to an expected-cost rule rather than as final answers.
What it does: For every ticket, one Jev call answers which team, how urgent, and whether the customer is about to leave. The router multiplies the probability of each possible mistake by what it would cost and sends the ticket to auto, review or human. It ships a dashboard, a review queue with an audit sample, Docker setup and a replay tool that compares routing policies without calling Jev.
Why it belongs in the showcase: A complete, tested triage service plus a reproducible calibration benchmark against three LLMs.
Caveats: Author-run benchmark; thresholds are illustrative.
Evidence
Public MIT README with CI; Choice (5 teams), Score (5-level urgency) and Noul (churn) in one request; cost routing with illustrative dollar thresholds or a simple 0.85 threshold mode; author's Banking77 benchmark (462 questions via OpenRouter): Jev 1.13 79.2% accuracy, ECE 10.4%, p50 346 ms, $0.07 per 1k vs GPT-4.1 mini, Claude Haiku 4.5 and Gemini 2.5 Flash; full run cost $1.60; runs offline by default.
Caveats
- Benchmark is the author's own run; accuracy gaps of 2–4 points across models are within noise by the author's own reading
- Jev measured 79% accuracy, below the 87% TypeSafe publishes for Banking77; a second synthetic-ticket run gave every model only 27–51%
- Cost thresholds in the code are illustrative and should be set per team
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.