.png&w=256&q=75)
🤓 Latest Submissions
.png&w=828&q=75)
Dispatch
Most routing agents spend tokens deciding how to spend tokens. Ours doesn't: routing is a local forward pass that never touches the API, so every routing decision costs exactly zero tokens under Track 1 scoring. At the core is a fine-tuned DistilBERT categorizer (ONNX INT8, 67 MB, ~11 ms per decision) that classifies each task into one of the 8 hackathon categories. A measured policy then maps each category to the cheapest engine that can actually answer it, walking a local-first cascade: deterministic solvers compute exact answers for math (0 tokens), a local Qwen2.5-3B handles the factual lane (0 tokens, 8.4 s/request on the judge's AMD Zen 4 hardware), and only what remains escalates to Kimi on Fireworks — the single billed step. Why fine-tune instead of heuristics? Regex hit 114/116 on validation, but only by matching verbatim cue words — on a cue-free paraphrase slice it collapsed to 32%, while DistilBERT held 100% (116/116 overall). Low-confidence predictions on token-saving lanes fail safe to the strong model, and generated code is verified by sandboxed test execution before it's trusted. The policy isn't hand-tuned either: an exhaustive per-category sweep against a pre-labeled answer cache picks the Pareto winner on accuracy vs. tokens, with zero new API calls. Official 10-task validation: 10/10 accuracy at 2,519 tokens — 26% under always-Kimi (3,415) and 35% under always-cheap (3,851; the "cheap" model is priced lower per token but ~1.7× more verbose, so it loses on both axes). Shipped as a Docker container meeting the full judging contract at 2 vCPU / 4 GiB.
13 Jul 2026
