01
Learned compact router
Waiting
P(escalate) —latency —
0 billed routing tokens
AMD Developer Hackathon · Track 1
Route locally.
Spend only when needed.
Watch a learned router make a zero-token decision in this browser, compare it with a paid prompt-based decision, and inspect provider-reported Fireworks usage in live mode.
18/19previous accuracy
97.14%v11 projected accuracy
9,685previous API tokens
1,824v11 expected API tokens
0v11 routing tokens
130/130tests passing
Live routing lab
See the decision cost.
Each run reveals the local route first, then performs the paid baseline classification, and finally returns an answer with exact provider usage when live.
Browser router loadedhashed-logistic-v1 · compact-80a9ce84c395 · 167 training labels · held-out 77.5% acc / 100.0% escalation recall
ZERO FIREWORKS TOKENS02
Prompt-based baseline
Waiting
The paid comparison starts after the local decision.
prompt N/Acompletion N/Atotal N/A
Waiting for baseline—
03
Answer stage
WAITING
The answer appears after both routing decisions are visible.
prompt N/Acompletion N/Atotal N/A
Answer model—
Per-query policy comparison
Run a query to compare
Both bars share the same answer usage, isolating the cost of paying an LLM to route.
N/Afewer policy tokens
Session totals
0queries
0learned policy
0prompt baseline
0tokens avoided
QueryLocalBaselineSaved
Your narrated walkthrough log will build here.
Submission architecture
Learn once. Route free.
The router is trained on empirical local-model outcomes. At judge time it spends zero Fireworks tokens and keeps deterministic verification as a second quality gate.
Score locally
The learned classifier estimates whether the local pipeline will succeed.
→
Generate locally
A quantized Qwen model handles the zero-token path inside the container.
→
Verify
Arithmetic, code, logic, and output constraints are checked deterministically.
↗only if needed
Escalate
An allowed Fireworks model repairs the small subset that needs more capability.
Evidence before resubmission
Measured under the real limits.
The new image is accepted only after a disjoint accuracy check, 2-vCPU timing rehearsal, router class metrics, and provider-reported token comparison against the previous 9,685-token run.
Inspect the benchmark code →