Axiom — Token-Efficient General-Purpose AI Agent

Axiom is a token-efficient general-purpose AI agent built for AMD Developer Hackathon ACT II — Track 1, where fewest tokens wins subject to a ≥84.2% accuracy gate. The architecture applies nine levers in combination: (1) thinking={"type":"disabled"} on every Fireworks call to skip the model's billed reasoning pass — the single largest token win; (2) tight per-category max_tokens caps (12 for math, 15 for sentiment, up to 180 for code generation); (3) stop sequences that halt on the first newline or follow-up cue; (4) JSON response_format for sentiment, NER, and logic to eliminate parse failures; (5) bounded SHA1-keyed in-memory prompt cache for identical prompts; (6) input truncation (200 chars for sentiment, 320 for summaries, 600 for code); (7) deterministic temperature 0.0; (8) ThreadPoolExecutor parallel processing at 4 workers; (9) shape validation pass that flags malformed outputs as INVALID before they reach the grader. On the 19-task AMD sample suite (fact_001–gen_003), the agent returns clean answers for all 19 tasks with 100% accuracy on the 12 tasks with verifiable ground truth. End-to-end token usage: ~1,545 total tokens (down from my previous 2,656). Wall-clock under 10 seconds even at single-worker sequential execution. The runtime is a single python -m app.main invocation wrapped in a slim python:3.11-slim container exposing port 8080 for health probes. The image reads /input/tasks.json, processes it through eight category handlers, and writes structured results + per-category token accounting to /output/results.json in strict conformance with the published schema. Single-model deployment keeps 404 risk at zero; model tiering is plumbed for the day cheaper chat models become available on the team's Fireworks account. See DESIGN.md for the full architecture narrative and scripts/token_report.py for a one-command local dry-run that validates accuracy, token budget, and zero-error conditions.