[Feature Request] Improve multilingual support by using local LLM for intent classification

[overdoignism]

First of all, thank you for developing ClawRouter. It is a fantastic project with great potential!

Currently, I've noticed that the conversation weight mechanism appears to be less effective when dealing with languages other than English. The static weight approach struggles to capture the nuances of different linguistic structures, leading to suboptimal routing.

Proposed Suggestion
I suggest introducing a pre-processing step using a local lightweight LLM (SLM) to evaluate the conversation before routing.

The workflow would look like this:

The incoming prompt is sent to a local model (By LM Studio or ollama etc).

The model analyzes the intent and the required capabilities (e.g., coding, creative writing, logical reasoning).

ClawRouter then routes the request based on a Capability-to-LLM mapping table.

Benefits
True Multilingual Support: Since LLMs are naturally better at understanding various languages than static keyword/weight rules, this would instantly improve the experience for non-English users.

Easier Maintenance: Developers would only need to maintain a mapping of "Capabilities" to specific LLMs, rather than managing a complex, hard-coded weight table for every language.

Scalability: New models can be added to the ecosystem simply by tagging their strengths.