Overfitters_General_Purpose_AI_Agent

Our Track 1 submission is a general-purpose AI agent built to handle eight distinct capability categories — factual knowledge, mathematical reasoning, sentiment classification, summarisation, named entity recognition, code debugging, logical reasoning, and code generation — through a single lightweight orchestrator rather than a heavyweight multi-agent framework. The pipeline works in four stages. First, a rule-based classifier assigns each incoming task to its category using deterministic keyword and structural heuristics, at zero extra token cost. Second, a model router maps each category to the right-sized model from the five allowed Fireworks models — cheap models handle sentiment and NER, a code-specialized model handles all debugging and generation tasks, and the largest model is reserved for genuine multi-step math and logical reasoning. Third, category-specific prompt templates keep every call lean, explicit about output format, and free of unnecessary chain-of-thought. Fourth, a validation and retry layer checks every response for malformed JSON, empty answers, and leaked reasoning before it's trusted, with a bounded retry policy and a guaranteed fallback so every task_id receives a non-empty answer even under total API failure. Every design decision was empirically tuned against real test logs rather than assumptions — including catching and fixing an on-demand model cold-start issue, reasoning leakage from a code-specialized model, wasted retries on non-retryable errors, and duplicate task_id handling. The result is an agent designed around how the competition is actually scored: pass the LLM-Judge accuracy gate first, then compete on raw token efficiency — not by cutting corners on correctness, but by never spending more tokens than a task genuinely needs.