Team 11
Surgical Vision
Run visual models once. Reuse their outputs.
A surgical review demo from Serelora and OpenSurgery AI that ties a depth orbit, a source-based report and cited clinical reasoning to one frame, and reuses model outputs between views.
Serelora and OpenSurgery AI
50 s
depth orbit
720p at 24 frames per second.
0 GPUs
for the render job
4 minutes 39 seconds on 4 CPUs and 8 GiB.
0
new clinical calls
A continuation reused the cached clinical reasoning response.
Figures are the team's own measurements from the hackathon weekend, as reported in their submission.
The problem
Rerunning visual models for each new camera move or report layout costs compute without adding evidence. Reusable, source-linked outputs keep review grounded.
What they built
A surgical demo that ties a depth orbit, a seven-page source-based report and cited clinical reasoning to one frame.
- Fuzzball execution: pretrained SAM3 and DA3 ran on a GB10, and the saved outputs then fed a CPU render and report job with no GPU.
- Selective recompute: camera moves and report layout revisions reuse masks and depth.
- Exact clinical cache: a continuation reused the completed clinical reasoning response.
- Intent routing: deterministic rules keep combined requests together and select supported handlers.
What is proposed
The team separates what it demonstrated from what it proposes. A VAST Data layer would store and search source video, masks, depth, metadata and vectors as reusable evidence objects. An upload event would trigger an adapter that checks source identity, model version and review state before requesting the right Fuzzball work. That adapter is not built yet.
Scope
The demo covers one unreviewed frame. Labels remain proposals and clinician review is required. Billed costs and realized savings were not measured.