🤝 Top Collaborators
.png&w=640&q=75)
🤓 Latest Submissions
Hybrid Token-Efficient Routing Agent
Track 1 asks for a general-purpose agent that handles eight task categories — factual Q&A, math, logic puzzles, sentiment, summarization, named-entity recognition, code generation, and code debugging — scored first on an accuracy gate, then ranked by token efficiency. Crucially, only tokens routed through Fireworks are billed. Our agent turns that scoring rule into its architecture. Every prompt is classified in microseconds by zero-token heuristics (regex/keyword rules, with no router LLM call), then answered on the cheapest backend that is still accurate: - Sentiment and summarization run on a bundled ~3B 4-bit local model (Qwen2.5-3B, via llama.cpp, CPU-only), so these answers cost zero Fireworks tokens — the best possible ranking outcome. - Accuracy-critical categories (math, logic, code, factual, NER) route to models chosen at runtime from the injected ALLOWED_MODELS list — nothing is hardcoded — each with a terse, category-specific prompt. Robustness is built in: local tasks run on CPU while Fireworks tasks run concurrently over the network, and a Fireworks fallback plus a deadline guard ensure the agent never emits an empty answer or exceeds the 10-minute wall clock. Output is always valid JSON with one entry per task. The container is self-contained (weights bundled, no GPU, no Ollama), 2.36GB, and was validated end-to-end under the grading box's exact limits (4GB RAM, 2 vCPU): all 8 categories correct, with sentiment and summary billed at zero tokens.
12 Jul 2026
