README.md

metadata

title: AMD Routing Agent Demo
emoji: 🧭
colorFrom: red
colorTo: gray
sdk: docker
app_port: 7860
pinned: false
short_description: Token-efficient tiered LLM routing cascade — live demo

AMD Routing Agent — Live Demo

This Space runs the judge-facing dashboard (routing_agent.webapp) for the Hybrid Token-Efficient Routing Agent built for AMD Hackathon ACT II, Track 1. Submit a prompt and watch it get routed through the Tier 0 (deterministic solver, zero tokens) → Tier 1 (cheapest adequate model) → Tier 2 (escalation) cascade, with live token/cost/savings reporting.

Main repository (source of truth, full README, PLAN, SECURITY audit): https://github.com/SebAustin/amd-routing-agent

This file is packaging only. See /Users/sebastienhenry/Documents/Hackathons/AMD Dev Hackathon/DEPLOYMENT.md in the main repo for how this Space is assembled and deployed.

---

How the orchestrator assembles this Space repo

This spaces/README.md and Dockerfile.spaces live in the main GitHub repo but are not used as-is by Hugging Face — the Space is a separate git repo that the orchestrator assembles from selected files out of the main repo. This keeps one source of truth (GitHub) while producing a Spaces-compatible layout (HF requires README.md + Dockerfile at the Space repo root).

Assembly steps (performed by the orchestrator, not by this file):

1. Create/clone the Space repo SebAustin/amd-routing-agent-demo.
2. Copy Dockerfile.spaces from the main repo root → Space repo root, renamed to Dockerfile.
3. Copy spaces/README.md (this file) → Space repo root, renamed to README.md (the YAML front-matter above is what HF parses for the Space card — title/emoji/sdk/app_port).
4. Copy src/ (the routing_agent package) → Space repo src/.
5. Copy evals/policies/ (routing policy YAML only — not the full evals/evalset/ corpus or evals/reports/, to keep the Space image lean) → Space repo evals/policies/.
6. Copy pyproject.toml, uv.lock, and the main repo's README.md (the package's own readme, required by hatchling at build time — see Dockerfile.spaces comments) → Space repo root.
7. Set Space secrets (see main repo DEPLOYMENT.md for the exact gated command — this step is never performed by copying a file; it is a hf CLI / web UI action against the Space's own secret store).
8. Push the assembled Space repo to the Hugging Face remote.

The harness/scoring Dockerfile at the main repo root is untouched by any of this — it stays the default docker run entry point for the scoring harness (ENTRYPOINT ["python", "-m", "routing_agent.adapter"]). Dockerfile.spaces is only ever used for the Space variant described above.