User repository on GitHub
Distill a repeated Jev classification task into a local model, on the fly — same answers, your hardware.
- Forks
- 11
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
Jev directory
CommunityJev Tools & CLIs
Jevstiller
A drop-in proxy that distills a repeated Jev classification call into a local model as it runs: it learns from Jev's own probability distributions, then answers most requests itself while uncertain ones and a 2% audit slice keep going to Jev.
- Stars
- 69
- Language
- Python
- License
- APACHE-2.0
- Reading time
- 2 min
- Published by
- Community
- Author
- tomerglick57
- Last verified
- Oct 2026
Highlights
- Runs next to your service as a proxy: docker run on port 8080, then point TYPESAFE_BASE_URL at it.
- Trains from Jev's answers with full probability distributions and answers in about 15 ms p50 on CPU against Jev's ~300 ms.
- Banking77: 71.9% of held-out messages answered locally at 99.50% agreement with Jev, against a 98% target.
- Uncertain and novel input plus a permanent 2% random audit slice keep going to Jev; a broken audit falls back automatically.
- One CPU process answers about 130 messages/s and a GPU about 2,000/s; the live 2026-09-25 run gave 70.7% at 99.45%.
Watch out
Apache-2.0. It distills one repeated classification task and needs a container-style deployment plus a TypeSafe key for the Jev fallback; the 71.9% local coverage on Banking77 is a replay against recorded answers, and on the tweets the system still matches Jev's modest 64% accuracy.
Questions about this entry
Short answers drawn from the linked project or page.
The routing threshold is chosen on a held-out, IID calibration set with an exact finite-sample Clopper-Pearson bound, testing candidates strictest-first, so the share answered locally and out of step with Jev stays inside your disagreement budget with 95% confidence.
One number, for example 0.98, becomes a 2% disagreement budget: Jevstiller returns the label Jev would have returned on at least that share of requests, and a permanent audit channel re-verifies the contract.
Everything. At first every request still goes to Jev; the proxy copies answers with their distributions, and the bundled recording shows the local model taking over after about 4,000 of 5,000 replayed customer messages.
No. If the audit shows the agreement contract is broken, Jevstiller falls everything back to Jev automatically, and the 2% audit channel keeps sampling traffic either way.
Anywhere Docker runs: ghcr.io/tomerglick57/jevstiller with a data volume. A single CPU process handles about 130 messages a second, your service keeps its own TYPESAFE_API_KEY, and local answers take roughly 15 ms p50.
Five public tasks: Banking77 (71.9% local, 78.5% accuracy for both Jev and the system), CLINC150, AG News, and TweetEval sentiment and offensive. On the tweets the student answers only the confident quarter and system accuracy still matches Jev's.
No. It distills one repeated classification task on your traffic; uncertain, novel and out-of-distribution input and the audit slice keep going to Jev, and the README targets workloads like agent loops and classify-then-act pipelines.
More Jev resources like this
A local workbench that pairs OpenCV vision with Laya or Jev decision models: images are segmented and measured, hard rules settle the numbers, and the model gives the verdict — with one-click installers and an "open-source JEV" training path served at /v1/systemone.
Jev Tools & CLIs#community#github#python
Community
uu889/laya-opencvPython
uu889/laya-opencvPythonToolkit, knowledge base and local dashboard for building with Jev via OpenRouter Alpha Decisions or the TypeSafe Direct API: a zero-dependency SQLite request log, a localhost REST server and Python CLI demos.
Jev Tools & CLIs#community#catalog-import#logicrw
Community
rileycarney/jevtoolsHTML
rileycarney/jevtoolsHTMLA Jev-style one-pass skill router for OmO agents: a frozen Qwen2.5-0.5B with a tiny option-attention head scores the whole skill catalog per prompt, and an extension shrinks the system prompt to the top-K skills, failing open.
Jev Tools & CLIs#community#routing#skills
Community
islee23520/omo-jevlike-routerPython
islee23520/omo-jevlike-routerPythonOllaya's site for its local decision-model runtime: one binary pulls open checkpoints and serves them on TypeSafe-compatible endpoints, now at Ollaya 0.8.0 with an app and command line for macOS, Windows, Linux, and Docker.
Jev Tools & CLIs#community#local#runtime
Community
jmurray10's MIT-licensed FastAPI service and library that classify business descriptions into one of 1,012 NAICS 2022 industries or hold them for a person, with Postgres search and code-owned gates doing the work and Jev only picking from a closed list.
Jev Tools & CLIs#community#naics#classification
Community
jmurray10/jev_naicsPython
jmurray10/jev_naicsPythonA logistics exception gate that reads carrier codes and free-text driver notes in one scan event and decides whether a shipment needs a human, which exception class it is, and how severe the delay looks.
Jev Tools & CLIs#community#python#logistics
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.