.png&w=256&q=75)
🤓 Latest Submissions
TRIARC
TRIARC is a three-tier, token-efficient AI agent built for the AMD Developer Hackathon Track 1: Hybrid Token-Efficient Routing Agent. Rather than sending every task to the most powerful (and most expensive) model, TRIARC routes intelligently. A local Tier-1 router first classifies each incoming task by the capability it actually needs — extraction, classification, simple code generation, complex reasoning, debugging, research, or synthesis. The model registry then resolves that capability to the cheapest Fireworks-hosted endpoint that satisfies it: a mid-tier Gemma model for moderate tasks, or a frontier model reserved for genuinely hard ones. If a worker's response comes back with low confidence, TRIARC doesn't just accept it — it escalates the same task to a stronger, costlier endpoint and retries, rather than defaulting every task to the most expensive model out of caution. This keeps token spend proportional to actual task difficulty. The system runs as a self-contained batch container: it reads a set of tasks from /input/tasks.json, routes and executes each one through Fireworks AI, and writes structured results to /output/results.json — no server, no manual intervention, ready for automated evaluation. TRIARC also includes an optional local management API and terminal dashboard (Textual TUI) for monitoring runs, cost savings, and routing decisions during development — though the core submission scoring runs purely through the batch container against Fireworks AI.
13 Jul 2026
