I am Kaman Singh Anand, my email is [email protected], contact me at 9604090790, please help me with a good group as i want to fuel my passion as a 16 year old, i would really apprieciate your help, i made this project all by myself and ai, i have no team so this was a solo, this is my first time participating in a hackathon and i could only make the track one project as given, i made it in 1 day, due to some critical difficulties i could not get more time, anyways here is my actual discription A cost-aware AI agent built for Track 1 that routes every question to the cheapest backend that can still answer it correctly, instead of sending everything through Fireworks. How it works: a free, instant classifier sorts each question into one of the 8 required categories. Factual knowledge, sentiment, summarization, and named entity recognition go to a quantized Qwen2.5-3B model bundled in the container, running fully offline at zero token cost. Math, logic puzzles, code debugging, and code generation route to Fireworks, since testing showed the local model reliably struggles with multi-step reasoning even with careful prompting. Built-in safeguards, found through real testing: low-confidence classifications get a free local second opinion before routing; local answers that look like non-answers auto-escalate to Fireworks; multi-part questions route straight to Fireworks after the local model was caught dropping the second half of compound questions; summaries with explicit length limits are deterministically trimmed to comply. Fireworks calls run concurrently, not one-by-one, keeping runtime well inside the 10-minute limit even on larger task sets. The container falls back gracefully at every failure point rather than crashing, and a final validation pass guarantees schema-valid JSON output. Built in Python with llama-cpp-python for local inference, Docker (linux/amd64), ~2.3GB compressed.
Category tags: