Leak Sight AI Workflow

Detection is a sequence, not a single frame.

Leak Sight AI narrows a visual change into a reviewable incident, then keeps an operator in control of the maintenance handoff.

Request a PilotReview the Sequence
RGB videoThermalOptional sensors
Temporal verificationEvent persists at Valve V-204Visual, thermal, and pressure context aligned
Incident ready for operator reviewHigh · 92% model confidence

Detection Sequence

Five stages turn a change into an action.

Each stage keeps the observation tied to a camera, monitored zone, asset, and time window. That traceability is what makes the output useful to a reliability team.

  1. 01

    Ingest the available evidence

    Combine supported RGB and thermal streams with optional pressure, flow, acoustic, weather, and asset context.

    Inputs stay tied to a known camera, asset, and monitored zone.
  2. 02

    Detect a visual change

    Evaluate dripping, pooling, steam, thermal differences, staining, corrosion, and damaged insulation.

    The supported scope covers visible water and steam around exposed equipment.
  3. 03

    Verify it over time

    Compare the event with recent frames and operating conditions so cleaning, condensation, shadows, and exhaust are less likely to become alarms.

    Facility calibration and temporal evidence matter as much as a single frame.
  4. 04

    Create a prioritized incident

    Package the probable source, location, severity, confidence, evidence clips, and recommended response for operator review.

    Operators can confirm or dismiss the event and preserve their reasoning.
  5. 05

    Move it into maintenance

    Export a confirmed event into the facility’s maintenance workflow with the evidence needed to assess and schedule the work.

    Confirmed incidents can be exported through integrations approved for the facility.
Operator turning a small red valve in a mechanical room with insulated pipeworkOperator Review

Evidence reaches the person responsible for the procedure.

Leak Sight AI can recommend an inspection step, but an operator confirms the event and follows the facility’s approved isolation and maintenance process.

Photo: Shixart1985 · CC BY 2.0

False-Signal Review

Look for persistence and context before escalation.

Cleaning, condensation, weather, shadows, and exhaust can resemble a leak. Facility calibration and temporal analysis are required because appearance alone is not enough.

Condensation or washdown

Check direction, duration, recurrence, and the monitored zone’s operating schedule.

Exhaust or ambient steam

Compare plume origin and motion with the known asset boundary and thermal region.

Weather and shadows

Use recent frames, site weather, and time-of-day context before escalation.

Expected process change

Align the visual event with pressure, flow, acoustic, or asset-state data when available.

Operator Handoff

The model proposes. The facility decides.

Critical actions remain under the facility’s approved procedure. Leak Sight AI packages evidence and a recommended next step; it does not autonomously isolate equipment.

  1. 01Review evidence

    Inspect the event clip, thermal region, signal context, and progression.

  2. 02Confirm or dismiss

    Record the operator decision and add facility context when needed.

  3. 03Route the work

    Export a confirmed event into the approved maintenance workflow.

Pilot Qualification

Test the sequence on a bounded set of exposed assets.

A pilot starts with known camera positions, high-risk locations, and an agreed operator review process.

Request a Pilot