Jev, distilled on the fly.

Put Jevstiller in front of a repeated Jev classification call. It learns from Jev's own answers, proves it agrees with Jev within a budget you set, and then answers most requests itself.

Get started View on GitHub

before   your app ──► Jev

after    your app ──► Jevstiller ─┬─► local model   most requests
                                  └─► Jev           unsure, novel, or 2% audit

Same call, same answers

You set one number, target_agreement. Jevstiller returns the label Jev would have returned on at least that share of requests, and keeps checking it on live traffic.

Past the rate limit

Jev caps out at 1,200 requests a minute. The local model runs at thousands per second on one GPU, with single-digit millisecond latency.

A bound, not a vibe

Routing thresholds are chosen against a Clopper–Pearson upper bound on held-out data, then re-verified forever on a random audit slice.

Falls back on its own

If the audit shows the contract is broken, every request goes back to Jev until a new model passes.

Why

Jev is fast, cheap and typed. It is also a ceiling: 1,200 requests per minute per key, ~350 ms per answer, hosted only. Jevstiller is for the workload that outgrows that: bursts, backlogs, latency budgets in milliseconds, boxes with no egress, or not wanting every classification to depend on one external API.

In an offline replay of Banking77 (77 intents), the local model took 71% of traffic at 99.5% agreement with the teacher, at ~2,000 rows/s on one GPU. How that was measured →

Status

Alpha (0.1.0). The loop is tested end-to-end and validated on replays with a perfect “oracle” teacher. The benchmark against live Jev is pending. The next milestone is a drop-in HTTP proxy: point TYPESAFE_BASE_URL at it and change no code.

Jevstiller is an independent open-source project (MIT). It is not affiliated with, endorsed by, or supported by TypeSafe.