v i m a.
Video intelligence for construction sites. No lidar.
Hardhat video becomes an inspectable evidence chain. Depth-delta filtering drops 57% of bad frames, MASt3R reconstruction grounds the rest, episodic memory binds events to time and zone.
evidence chain · one scroll, one proof path
every payout starts as a frame you can inspect.
vima turns egocentric video into timestamped work claims, then binds those claims to spatial zones before settlement logic sees them.
timestamped bodycam frames become an inspectable CII trail before anything touches payout logic.
frame stream · drag to inspect
episode taxonomy · five shapes from the paper
every claim is one of five shapes.
Vima's episodic memory emits one of five episode types per detected event. The shape makes the claim reviewable in seconds. A human swipes, the claim resolves, the reviewed evidence enters the audit trail.
masonry_work_candidate — observed work in progress on the masonry plane
scaffold_zone_visible — scaffolding plank or rail in the active frame
safety_edge_context — exposed edge or fall hazard in view
foreground_worker_present — primary subject in motion or labor
ledger · settlement receipt
this is where a frame becomes inspectable evidence.
the ledger is the audit handoff: frame labels stay visible while the payout gate scores productive time, blocks idle work, and preserves the receipt.
verify · human review of episodic memory claims
every claim earns its truth.
Centralized labeling pipelines move at a worker's reading speed. Vima reduces the unit of work to one tap on a phone. The model proposes its episodic memory claims, humans confirm or reject. Right to accept, left to reject, up to skip. The reviewed claims become the audit trail every safety incident depends on.
- · tinder-style swipe deck
- · accept · reject · skip per claim
- · episode types match the paper's five
- · human review feeds the audit trail
- · every reviewed frame stays inspectable
- · reviewer decisions become ground truth
pipeline
video in. auditable work claims out.
