Market Watch Commander

AI agents can write analysis code, but running that code introduces two challenges: incorrect results and uncontrolled execution. Market Watch Commander addresses both through a market investigation workflow built on Vultr. Users select a historical window to investigate Bitcoin, Ethereum, and Solana. The agent uses Vultr Serverless Inference to generate Python, executes it inside a disposable Docker container, and checks its numerical results against independent DuckDB calculations before explaining the evidence. A comparison heatmap helps users explore buying and selling activity across eight days. PostgreSQL preserves investigation history, execution events, evidence, and reports. Each sandbox has no network access or credentials, read-only inputs, resource limits, and an externally enforced 10-second deadline. Our confirmed containment demonstration deliberately introduces an infinite loop: the runner terminates and destroys the container, then allows one correction attempt in a fresh container. Vultr hosts the dashboard, backend orchestration, databases, and sandbox execution, while also providing agent inference. The project demonstrates how autonomous analysis can be useful, observable, and bounded, using historical market data as a practical use case.