Skip to Main Content

thermomind

Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

thermomind — Demo video

ThermoMind is an evidence-first AI investigator for gas-turbine power plants that pushes Meta's models into a physical system. It reads real turbine measurements (36,733 hourly rows from a public dataset), compares them with a measured-data baseline, and uses Meta Muse Spark with verifiable tools to investigate deviations: every number the model cites is traced to a row, a calculation and a model version, and a server-side validator rejects anything else. Around that loop sit plant-wide agents: findings route to the responsible engineer and supervisor, subsystem agents confer about the domino effect before anyone acts, a watchdog notices needs on its own, and a calibrated physics simulator lets teams rehearse faults on a model. An explicit autonomy policy says what agents may do alone and what only humans may do; the software has no write path to a control system. Startup track: the auditable case file, not the alarm, is the unit of value.

Team