LOCAL INFERENCE · PUBLIC EXPERIMENT
Small model.
Clear decisions.
A decision layer for the questions before the answer.
Try chatbot and agent routing in English or Vietnamese.
The decision
02 / RESULTOne request. A typed decision.
You’ll see the route, model score and response.
No tools are executed.
BUILT TO BE INSPECTED
A routing experiment,
with its limits in view.
Multilingual MiniLM encodes the request. Two small trained classifiers select a topic or suggested skill. A simple margin rule can ask for clarification; responses come from fixed templates.
The first Windows CPU run classified 11 of 12 chatbot requests and 7 of 8 agent requests correctly. These handwritten fixtures are small and not an independent benchmark. See the case study for separate LXC measurements and observed failures.
Architecture, measurements & failure cases →Your text stays out of application storage.
No chat history or request bodies are logged by this app. Public requests travel through Cloudflare before reaching the homelab. Please don’t submit sensitive information.
Recommendations don’t grant permissions.
The demo has no private documents, credentials or executable tools. Agent actions are simulations. This is an independent local baseline, not Jev or a general-purpose chatbot.