.png&w=256&q=75)
š¤ Latest Submissions
.png&w=828&q=75)
TokenCascade
TokenCascade is a token-efficient routing agent built for the AMD Developer Hackathon Track 1. Rather than classifying task difficulty upfront, it uses a cascade: every task is first attempted by a local Qwen2.5-1.5B-Instruct model running free on-device. A task-specific verifier then checks whether that answer can be trusted ā deterministic checks for math (including a zero-cost Python calculator for clean arithmetic expressions), format validation for structured/JSON output, exact matching for classification tasks, and self-consistency sampling for open-ended or ambiguous cases. Only when the verifier rejects the local answer does the agent escalate to Fireworks AI's gpt-oss-120b, the most cost-efficient model available on the platform for this use case. This design directly optimizes the scoring objective: correctness stays high because escalation is trustworthy and targeted, while token cost stays low because the local model handles the majority of tasks for free. The system runs as a portable Docker container reading /input/tasks.json and writing /output/results.json, with retry and timeout handling to stay robust under evaluation constraints.
13 Jul 2026
