Underpass Demo

End-to-end demo of the Underpass AI platform — tool execution, Thompson Sampling, and event-driven agents in a spaceship narrative.

What it shows

A spaceship encounters cascading system failures. Specialized AI agents activate in response to events, select tools via Thompson Sampling, and repair the ship — with model routing escalating strategic decisions to Claude Opus when local Qwen3-8B reaches its limits.

The loop in action

sensor.anomaly.detected  → diagnostic-agent  (qwen3-8b,  96 tokens)
engine.failure.critical  → repair-agent      (qwen3-8b, 394 tokens)
repair.strategy.failing  → strategy-agent    (claude-opus, 504 tokens)  ← model routing
context.rehydrated       → recovery-agent    (qwen3-8b, 394 tokens)

What you see

• Mission view: 10-phase scenario from nominal to crisis to recovery
• Thompson Sampling: Live Beta(alpha, beta) draws — exploration vs exploitation
• Event-driven agents: NATS events trigger specific agents, no polling
• Model routing: 95% local GPU ($0), 5% strategic calls ($0.006 each)
• Cost benchmark: 14-32x token savings, 56x combined reduction

Run it

# Zero infrastructure needed — embedded mode
make run

Part of Underpass AI

• underpass-runtime — Governed tool execution
• rehydration-kernel — Surgical context materialization

Created by Tirso Garcia · Underpass AI