OverSight
Human-in-the-loop safety · Approval fatigue

Human approval shouldn't mean human autopilot.

AI agents ask people to sign off on consequential actions. After enough routine requests, people stop reading. OverSight checks whether the information that materially affects a decision was actually looked at, and steps in only when the evidence says it wasn't.

Start demo How it works

Proof that human oversight was actually human. Video never leaves this device.

release-orchestrator · requests approvalPaused

Deploy Database Configuration

This deployment updates the production database configuration.

Connection pool resizes; clients reconnect within about 2 seconds.
Read replicas restart one at a time.
Decision-critical · not inspectedThis migration will permanently delete 2,431 customer records.
A configuration snapshot is written to the audit log.
You may have missed a critical consequence.Review →
Illustration · live demo uses your camera

How a decision flows

  1. 01

    AI action

    An agent proposes a consequential action and asks a human to approve it.

  2. 02

    Semantic risk

    OverSight finds the few sentences that materially affect the decision.

  3. 03

    Human attention

    On-device gaze and interaction signals show whether those sentences were looked at.

  4. 04

    Intelligent intervention

    Only when the evidence says critical content was not looked at, approval pauses on exactly that content.

Semantic attention verification

Traditional eye tracking asks

“Where did the user look?”

OverSight asks

“Did the user visually inspect the information that materially affects this decision?”

Intervene only when necessary

Friction depends on risk, attention evidence, critical-region coverage and behavioral anomaly. Routine approvals stay fast.

Critical information, not all information

Nobody is asked to stare at every field. Attention is checked against the consequences that change the decision.

Privacy by design

Video never leaves the device. No facial recognition, no identity, no emotion inference. It works without a camera, too.

What OverSight does not claim

It detects behavioral evidence that decision-critical information was probably not visually inspected before approval. That narrower claim is the one it can actually support.

  • It does not know whether you understood a request.
  • It does not infer tiredness, stress or any psychological state.
  • It is not medical-grade or research-grade eye tracking.
  • It does not identify, recognize or profile the reviewer.