Walkdown

Walkdown turns a plant walkdown video into a verified corrosion register and a repair plan that a person approves. The problem: corrosion costs industry about $2.5 trillion a year (NACE IMPACT), yet walkdown findings still end up in notes and spreadsheets, and repair plans are made by hand. How it works: an inspector films a walk through a plant. A Muse Spark agent decides what to inspect, SAM 3.1 segments and tracks possible corrosion, Muse Spark grades every finding with written evidence, and Muse Wonder gives a second opinion on uncertain or severe cases. Muse Voice Transcribe turns the inspector's voice notes into asset tags and line service. Every finding goes into a register where a person approves, rejects or sends it for re-inspection, choosing between the two model opinions when they disagree. An OR-Tools CP-SAT solver then schedules the repairs around crews, skills, permits, shutdown windows and risk, and replans in seconds when something changes, keeping started work fixed and showing the added risk. Nothing takes effect until a person approves the plan. A copilot answers questions from the register and plan only, and the app exports a PDF report plus CSV/JSON work orders. On-prem mode runs on open weights (SAM 3.1 on Modal, Muse Spark Open on Baseten/Fireworks) for sites that can't send footage off-site. Measured, not claimed: in a blind human audit the default setting kept 110 of 113 real rust findings (97.3%), 12% of 191 clean-pipe images still got a false flag, and severity grades are not yet validated, which is why a person reviews everything. 50/50 live end-to-end agent runs succeeded. Every credit dollar is itemised in docs/CREDIT_USAGE.pdf. Decision support only. Live demo: https://walkdown-zeta.vercel.app