system-one-categorizer-demo
Bulgarian blog categoriser comparing Jev and local Laya on 50 posts, with a small local LLM naming new topics.
Overview
This demo app, documented in Bulgarian with an English summary, does two things: compares two System One models on Bulgarian text and shows a working System One model inside an application. You write a blog post and the model suggests a category, either an existing one or 'new topic'; you accept or correct it with one click. When a new topic is needed, a small local generative model (BgGPT 4B via Ollama) names it and writes a short description, since a System One model does not write text. The human always decides.
On 50 labelled Bulgarian posts, Jev (jev-latest) was right 48/50 versus 36/50 for Laya multilingual running locally on CPU: 28/28 versus 24/28 on existing categories and 20/22 versus 12/22 on new topics. Jev was correct on all 44 answers with confidence at least 0.95 and flagged none as uncertain, while Laya flagged 17 below 0.5. Median latency was 380 ms via API versus 200 ms locally, and Jev cost about $0.001 for 50 posts. Jev can be reached via TypeSafe, DigitalOcean, OpenRouter or Vercel. MIT licensed.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which of the blog's categories does this post belong to, or is it a new topic? | — |
How it works
- User writes a post
- Jev or Laya chooses an existing category or 'new topic'
- If new, BgGPT names the category
- User accepts or corrects; new categories join the list
Strengths
- Clear side-by-side comparison on a non-English language
- Confidence shown and low confidence flagged
- Free local option with Laya
Caveats
- Only 50 evaluation posts
- Laya often forces new topics into existing categories
Getting started
uv pip install -r requirements.txt (and typesafe-sdk for Jev), put TYPESAFE_API_KEY in .env, then python app.py.At a glance
- Type
- App
- Status
- prototype
- Language
- Python
- Runtime
- Local web app
- Has code
- yes
- License
- MIT
- Stars
- 0
Use cases and integrations
How curated lists describe it
Jev vs Laya on Bulgarian: two System One models suggest a category for each blog post. A local BgGPT names new topics.
wh000wh000/awesome-jev-live
Frequently asked questions
Which model did better on Bulgarian?
Jev, 48/50 versus 36/50 for Laya multilingual.
Who names new categories?
A local BgGPT 4B model, because System One models do not write text.
Can it run without a key?
Yes, it falls back to local Laya.
Related projects
Library, CLI and app that classify PDFs, DOCX and PPTX with Jev and split document packets at predicted boundaries.
BenchmarkbetaPythonDocker recipes and ARM64 images for serving ten native decision checkpoints, such as Cloudflare Clef, behind /v1/systemone on DGX Spark.
ServicebetaPythonZero-dependency Node proxy that shadows Jev with local decision models and reports whether you can safely swap, per question and policy.
ProxybetaJavaScriptBrowser arena where local Laya and hosted Jev race in Snake and fight in a Mortal-Kombat-style game, with every move a real decision.
ServiceprototypePythonChrome T-Rex game where local Laya on MLX and hosted Jev choose moves side by side, with latency charts and crash replays.
ServiceprototypePythonBrowser Pac-Man race comparing hosted Jev and self-hosted Laya choosing routes from the same board state, after code removes fatal routes.
ServiceprototypeJavaScript