SurvAI

SurvAI puts one engineer on site instead of six. Before a router goes into a telecom equipment room, an engineer and five departments (Space/Acceptance, Power, Cooling, Execution, Room Manager) each give a working day to visit the room together, and the signed survey takes 3–5 days. With SurvAI, one engineer films the room; Meta models read the footage and one AI agent per department holds the meeting in minutes; the departments review the report and sign off from their desks, and a representative's objection goes back into the next meeting run. At the 1,000+ surveys a year one region runs, that is 5 of 6 site visits saved, about 5,000 person-days a year. How it works: the engineer uploads the walk-through as filmed. In the browser SurvAI picks the 30 sharpest frames and Meta SAM 3.1 finds the labels, taped sheets, stickers and people to blur; only blurred frames are stored. SAM 3.1 outlines racks and floor tiles, Muse Spark reads the Egyptian Arabic narration of the cable route and drafts the room (marking the rack the engineer named), and the engineer corrects the plan on a tile grid. Then the department agents meet for up to three rounds: each sees the others' statements, objects when a rule it owns is broken, cites the frames and calculations that prove it, and asks the engineer what the footage can't show. Muse Wonder leads as Survey Lead, resolves conflicts, places the device and draws cable paths (never through a door). Door clearance, breaker capacity, heat load, rack space and power/data separation are calculated in code and force an objection the model cannot soften. The engineer approves (an override needs a written reason), and the report mirrors the company's survey form with a floor plan to scale. Measured on the sample room (docs/eval/accuracy.md): datasheet 8 of 8 fields match Juniper's guide; SAM rack boxes 43 of 44 correct; taped sheets 9 of 9 found; the meeting picks the same rack in 5 of 5 runs (on a local test survey of the same room). Try it without a passcode: https://survai.ramy-comm.workers.dev/sample ("Run the meeting again, live" runs the real agents on the same room). In the sample, the departments agree on rack R4 and find the blocker before anyone travelled: the room has none of the two free 40 A DC feeds the router needs, so power is arranged before the install visit instead of being discovered on it. Judges: create your own survey with this link (5 a day): https://survai.ramy-comm.workers.dev/#judge=judge-4eeb5875d2a5 Test kit for that survey (one zip: the walk-through used in the demo with the engineer's Arabic narration, the power panel and rack photos, and a README with the steps): https://drive.google.com/file/d/1IHSdgxA3Ts2BzYcxgNG6uCFAOkmRhL8r/view