Marshal

Marshal is a situational-awareness agent for GPU data center operations, built on Crusoe Managed Inference. It constructs a live thermal model from streaming rack telemetry, predicts rack-level throttling about five minutes before it hits a specific rack, and surfaces one plain-language advisory a non-technical shift engineer can approve, question, or override in the moment. It is an agent, not a dashboard: the rack view is context, the proactive advisory is the product. Code computes every number (a first-order thermal model gives the five-minute forecast; the model never does arithmetic). The language model does what a rule cannot, shown on screen rather than claimed: it keeps a job co-located with its gradient partner instead of migrating to the emptiest rack, beside the flawed pick a headroom-only rule would make; and it interprets a free-text override, even one that names a rack only by description ("the rack running the checkpoint writer has a firmware update"), into a structured constraint it then learns from for every later decision. Advisory reasoning runs on NVIDIA Nemotron-3-Ultra-550B with a DeepSeek-V4-Flash triage tier via Crusoe Managed Inference, deployed live on Cloudflare Workers and Durable Objects. Built entirely during RAISE Summit Hackathon 2026 (Crusoe track, solo, remote). Rack telemetry is simulated; the agent, the inference, and the learning are real.