alloy-agent

alloy-agent is an AI agent for steel alloy design. You state the properties you need, in plain language or as a spec-sheet PDF or photo. A Muse Spark agent turns them into constraints, searches candidate compositions with an uncertainty-aware predictor, explores the strength-ductility trade-off, and returns ranked candidates with calibrated ±2σ bands and a recommendation of which alloys to test next. The problem: alloy design is a slow guess-test-repeat loop, and LLMs readily invent plausible-looking numbers. alloy-agent is built so it cannot do that. Every number in the answer must come from a tool output, and a code verifier strips anything it cannot trace. Out-of-domain candidates are flagged, never hidden. When a goal is out of reach under the conservative mu-2σ criterion, the agent says which constraint binds and by how much, then offers the Pareto front instead of inventing an alloy. Results: 20/20 task success on 20 evaluation queries with both Muse Spark Open (Baseten) and Muse Spark 1.3. The predictor's ±2σ bands cover 95% of held-out steels. Everything was built during the hackathon from an empty repository, on open data (MIT). Live demo: https://lee-ok--alloy-agent-ui.us-east.modal.run (Modal; the first request after idle takes about 30 s).