.png&w=256&q=75)
š¤ Latest Submissions
.png&w=828&q=75)
NeuroRoute
NeuroRoute is a general-purpose AI agent built for the AMD Developer Hackathon Track 1 challenge, designed to handle eight distinct capability categories: factual knowledge, mathematical reasoning, sentiment classification, text summarization, named entity recognition, code debugging, logical reasoning, and code generation. Rather than routing every task to a premium API, NeuroRoute uses a lightweight rule-based classifier to identify each task's category and intelligently decides whether a small local model (Qwen2.5-1.5B, running entirely on-device with zero token cost) can handle it reliably, or whether the task requires the added reasoning power of a Fireworks AI model. Simple tasks like sentiment classification, summarization, named entity recognition, and factual questions are answered locally at zero cost. Harder tasks involving multi-step math, logical deduction puzzles, and code debugging/generation are routed to Fireworks AI to preserve accuracy. This approach mirrors a real-world enterprise pattern: keep inference costs low by using in-house models wherever possible, and reserve costly premium API calls only for tasks that genuinely need them. The result is an agent that balances strong accuracy across all task types with meaningful token efficiency. The agent is packaged as a Docker container that reads tasks from a JSON file, processes them through the router, and writes structured JSON results, built for the linux/amd64 grading environment within the 4GB RAM / 2 vCPU constraints.
13 Jul 2026
