Community · E2 · artifact verified
Route messy music requests before the LLM
A DJ chatbot's pre-LLM router, built as a learning study: free-form /play requests - English, Spanish, typos, pasted lyrics, emojis - classified by Jev in one call (~500 ms, ~$0.000036) into a structured hint that tells the expensive LLM what it is handling.
01 · Role in the system
What Jev does here
The author's method was to learn by attacking: build a real router, then throw increasingly nasty requests at it. Each messy /play input becomes one typed classification call through the Vercel AI Gateway; the calibrated answer becomes a structured hint - request kind, what to watch out for - that front-loads the LLM's work, so the expensive model starts from a typed summary instead of raw chaos. The repo doubles as the lab notebook: a benchmark script replays adversarial request suites, and two dated result files archive what the routing achieved and where it broke.
02 · Control boundary
Where Jev sits
Jev as the pre-LLM classifier of a music chatbot: messy multilingual requests become one typed call each, calibrated answers are shaped into a structured hint, and the expensive LLM handles only the interpretation the hint frames.
Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.
03 · Known limits
What this evidence does not prove
- Calls route through the Vercel AI Gateway rather than the native endpoint.
- Two dated result files are author-run; no benchmark against a regex or classifier baseline.
- 0 stars; a personal learning study, not a maintained library.
04 · Attribution
Public sources
This is a Community record: the project was published by a third-party community author.
- panchicore ↗Community · github · public · checked 2026-10-07