This is Kev-0.8B fine-tuned for Nepali support routing. It does not write an answer. You give it a state (the customer's message) and typed questions; each question is one forward pass that scores every option you offered, so the result is a probability per option. It can't invent a category or return broken JSON.
- choice which team gets the ticket (11 categories per service)
- noul does the agent have to ask a follow-up first?
- 0 output tokens. Everything runs on this device.
State
Ctrl + Enter
Presets held-out test messages, never trained on
Ask your own questions (each one adds a forward pass)
The router was fine-tuned on the two questions above. Kev still answers others; how well depends on the question.
Decisions
Load the model, pick a preset or type a message, then Route.
request
response
As a smart if
Run every preset
Each message is routed one after another with both questions, the way a backend would call it in a loop.
| message | expected | predicted | conf. | clarify | ms |
|---|
How it works
- Every question becomes one token row:
<state> message <q> instruction <opt> option </opt> … <decide>. The instructions and option descriptions are English; only the message is Nepali. - The backbone (Qwen3.5-0.8B with the fine-tuned LoRA merged in, exported to ONNX) runs the row once. No tokens are generated.
- A pointer head compares the hidden state at
<decide>with the hidden state at every</opt>. A softmax over those scores, at the calibration temperature, gives the probabilities.