Abnu is a Dockerized general-purpose AI agent built for Track 1 of the AMD Developer Hackathon ACT II. On startup it reads a batch of natural language tasks from /input/tasks.json, solves each prompt, and writes valid answers to /output/results.json before exiting cleanly. The agent handles all eight required Track 1 capability areas: factual knowledge, mathematical reasoning, sentiment classification, text summarization, named entity recognition, code debugging, logical and deductive reasoning, and code generation. It uses lightweight per-task classification and category-specific handling to keep responses accurate and concise. Abnu follows the official Fireworks AI runtime contract. It does not hardcode API keys or model IDs; instead it reads FIREWORKS_API_KEY, FIREWORKS_BASE_URL, and ALLOWED_MODELS from the judging harness environment at runtime and routes all model inference through the provided base URL, using only permitted models. The container is resilient by design: it always writes valid JSON output, handles malformed or edge-case inputs without crashing, and exits with the correct status code within the runtime budget. The final image is a lightweight python:3.11-slim build for linux/amd64, publicly pullable and ready for automated evaluation. Abnu prioritizes clearing the accuracy gate reliably, then minimizing total tokens through efficient routing and compact, structured output.
Category tags: