.png&w=256&q=75)
🤓 Latest Submissions
Token-Router: Hybrid Token-Efficient Routing Agent
Token-Router solves the core tension in Track 1: local models are cheap but unreliable, remote models are accurate but token-metered. Rather than picking one, the agent routes every task individually down a ladder of increasingly expensive checks, exiting as early as it safely can. For each task: the local model answers first. A deterministic, zero-token judge checks the answer against task-specific rules (valid JSON? contains a number? within a word limit?) — when a rule applies, the decision is free and immediate. When no rule applies, the agent falls back to self-consistency: it samples the local model a few more times and checks whether it agrees with itself. Agreement signals confidence and the answer is committed locally at low cost; disagreement signals the model is guessing, and the task escalates to the remote model. Only tasks that genuinely need it ever reach Fireworks. The system is built around a single swappable model interface, so the same routing logic runs unchanged against mock models, a local Ollama server, or Fireworks, launch-day model swaps are a one-line config edit, not a code change. It includes offline tooling to grade committed answers against gold references and sweep the self-consistency threshold, producing an accuracy-vs-tokens frontier so the cheapest setting that clears the accuracy bar can be read straight off a table. The whole pipeline is resilient by design: an unreachable local endpoint forces escalation rather than crashing, and a failed remote call falls back to the local answer, so no single failure can zero out a run. Fully containerized with a documented, runnable README and MIT license, validated end-to-end against real local (Ollama) and cloud (Fireworks-compatible) models before submission.
13 Jul 2026
