Skip to content
Gray Scrubs AI

00:00

One model holdsthe state of theoperating room.

Fifty seconds of simulated surgery on a phantom, reconstructed from ordinary cameras with no depth sensor in the room.

Talk to a founder

Programmes

  • Inception
  • Activate
  • Claude for Startups
  • NIC Karachi

One signal, five readings

One signal, five readings

Surgical phase

under development

Sterile technique

open research

Traffic and turnover

5 roles present

Care-bundle events

open research

Metric 3D twin

42,483 points

Roles in this room

The most expensive thing that goes wrong in an operating room is still audited by hand.

A quarter, audited by handA file of every case in a quarter, hung on a rail, with two empty slots. One thin trace leaves the file only at those two slots, runs through the two spot-check checklists and ends at a single clipboard tally counted by hand. Every other case is never looked at.
An illustrative quarter: every case in it, the one or two a compliance spot-check looks at, and the clipboard they are counted on by hand. A diagram, not model output
3-5%
Surgical site infection rate after orthopaedic procedures
CDC NHSN 2024
$28K
Average direct medical cost of a single infection
AJIC orthopedic SSI review, 2023
68%
Infections not identified until after discharge
CDC NHSN 2024
1-2
Compliance spot-checks a quarter, counted by hand on clipboards
Current standard of practice

A room is not a list of detectors.

Gray-1 in the room: ceiling cameras, one device, five readoutsAn isometric operating-room installation on one deck. A ceiling camera rig on four posts carries four ordinary cameras angled down at an operating table. Their lines run down one conduit into a Jetson-class device on a low stand, which holds the one description of the room. One trunk leaves the device and branches into five identical reading plates on a rail: surgical phase, sterile technique, traffic and turnover, care-bundle events and the metric 3D twin.01 CAMERAS02 ONE DESCRIPTION03 READOUTS
An illustrative Gray-1 room: ceiling cameras, the Jetson-class device every model runs on, and the five readouts read off the one description it holds. A diagram, not model output

The drawing shows cameras in the room.

Why we do not quote the field’s sterility number

Five readouts, one model underneath them.

One description, five readingsAn exploded stack read from the top: five thin plates over one base slab. The base carries a plan of an operating room traced from our own reconstruction of it: the walls, the doorway and the two tables. Each plate above is a reading registered to that same room: surgical phase as a time band in four segments with a marked head, sterile technique as the field outline over the table with one contact mark at its edge, traffic and turnover as the doorway with entry ticks and an interval between two cases, care-bundle events as checklist rows on a shared time axis, and the metric 3D twin as the real people standing as points, with one measured distance between the two tables. One shaft rises from the base through a cut-out in every plate.SURGICAL PHASESTERILE TECHNIQUETRAFFIC AND TURNOVERCARE-BUNDLE EVENTSMETRIC 3D TWINONE DESCRIPTION
Five readings over one room. The base is our reconstruction of a public dataset of simulated surgery on a phantom, seen from above; the plates over it, surgical phase, sterile technique, traffic and turnover, care-bundle events and the metric 3D twin, are illustrative. Base: model output. Plates: a diagram

  1. Surgical phase

    under development

    Decided once per second while the case runs, and developed against a public dataset of simulated surgery on a phantom. No clinical footage yet.

    Why the decode commits

  2. Sterile technique

    open research

    A breach and the phase it happened in come from the same description of the room, so the two can never contradict each other.

    Why that number is a rule

  3. Traffic and turnover

    open research

    Every additional entry during a sterile phase degrades operating-room air, which is the documented mechanism linking traffic to infection risk.

    How an entry is counted

  4. Care-bundle events

    open research

    The checklist comes from our own clinical team and follows the prevention bundle a theatre already works to. No outcome data yet, on any site.

    How the taxonomy was built

  5. Metric 3D twin

    open research

    Cameras only, with no depth sensor in the room. Distances are checked against a laser measurement of the same scene.

    How the twin is built
How Gray-1 works

Everything we find, dated, including the results that went against us.

All notes
24 August 2026ResearchA metric twin of the room, from cameras aloneMulti-view RGB in, a metric 3D reconstruction out, with every point carrying an entity label and pairwise distances validated against time-of-flight ground truth.17 August 2026EvaluationThe headline sterility number is a rule, not a modelThe 0.768 that the operating-room literature quotes for sterility-breach detection is a zero-shot proximity heuristic, reported without an alarm rate. We read it at source.04 August 2026MethodPhase detection that has to commitA causal, forward-only decode at one inference per operating-room second. A model that cannot revise its own history is harder to build and is the only kind that is useful in the room.

We are looking for the operating rooms to be wrong in.

Talk to a founderHow a pilot runs