Classify
Route each prompt into one of eight specialized capability paths.
AMD Developer Hackathon · Track 1
Reason locally.
Escalate intelligently.
A general-purpose AI agent that protects answer quality while minimizing metered model tokens through local inference, deterministic verification, and adaptive routing.
94.7%acceptance score
8capability domains
49s19-task run
54/54tests passing
Interactive demo
One agent. Eight kinds of work.
Choose a capability, edit the prompt, and see how the routing layer responds.
System design
Accuracy first. Tokens second.
The router only trusts answers it can verify, keeping local inference useful without gambling on the accuracy gate.
→
Generate locally
Run a quantized 3B model inside the container at zero scored tokens.
→
Verify
Check arithmetic, code, constraints, entities, and confidence deterministically.
↗unverified only
Escalate
Call an allowed Fireworks model only when quality needs reinforcement.
Built for the harnessNot a prototype.
Not a prototype.
A measured system.
The same container contract used by the evaluator: strict resource limits, atomic output, runtime-only credentials, and adaptive pacing.
Inspect the public image →