khanuzeb/axiom

AXIOM is a system that solves math, physics, and chemistry problems using multiple AI agents working together. The idea is that a single AI asked to do everything at once, symbolic math, code, research, and plotting, tends to do all of it poorly. AXIOM addresses this by splitting the work among specialist agents that run in parallel, then checking the results before returning an answer. When you submit a question, a coordinator agent decides which specialists are needed and sends the query to them simultaneously. Specialists cover areas like math and physics, reasoning, code, research, visualization, chemistry, and retrieval from uploaded documents. Their outputs go to a critic that operates in two layers. First, it programmatically re-derives the algebra using a symbolic math library, cross-checks numeric and symbolic answers, and validates units. Second, an LLM-based audit looks for contradictions between agents, missing reasoning steps, and invented formulas. If the critic finds problems, it sends the work back with suggestions, up to a configurable hop limit. An aggregator then combines everything into a final response. Shared state is designed so parallel agents can write results without overwriting each other. The system also includes memory for past interactions and retrieval from uploaded PDFs. Both Qdrant and Redis are supported for memory and caching, but neither is required. If they are not running, AXIOM falls back to local alternatives. The project provides a FastAPI backend with endpoints for solving problems, streaming progress, uploading documents, and document Q&A. A Streamlit frontend offers a simpler form-based interface. Simple algebra questions skip the full agent pipeline and use a deterministic solver instead, which means they are returned without the critic's review. Known limitations include code execution that relies on a subprocess blocklist rather than a hardened sandbox, so untrusted code should not be run through it. The critic's quality depends on the underlying LLM, and there is no token-level streaming. The full README is longer than what was shown.