The wiseDo Platform · component 01
wiseDo Sight: what happened, and proof it happened.
Sight puts AI on the IP cameras you already have. It finds every person on the floor, holds one identity across cameras and occlusion, and checks the rules your site actually runs on. Every finding lands in a sealed, hash-verifiable record with a corrective-action workflow behind it, so what reaches an auditor is evidence rather than a screenshot. Processing happens on an edge box inside your plant.
How it works
See. Check. Record.
A closed loop on every camera, running at the edge inside your plant. We store events, not video.
See
A detector finds every person and the protective equipment they are wearing; a re-identification model keeps one person as one person across occlusion and across cameras.
Check
Zones and rules you wrote, reviewed and released are evaluated against what the cameras saw. The rule vocabulary is closed on purpose: there is no expression language for a mistake to hide in.
Record
Findings reach the supervisor responsible for that zone, and the evidence lands in a sealed bundle a third party can verify without us. Repeat findings open a corrective action.
What Sight does today
The capabilities, with their real readiness.
We publish what is running and what is not, because the gap between the two is where trust is lost. Sight ships as a safety and compliance evidence system first; process measurement is the next layer on the same perception stack.
PPE compliance
Helmet and high-visibility vest, detected per person per frame. Five served classes, nothing inferred from a sixth.
Restricted-zone entry
Zones drawn on the camera view, with entry evaluated against who is permitted and under what conditions.
Tracking & re-identification
One person stays one person through occlusion and across camera handover, so a rule is applied once rather than once per view.
Governed zone rules
A rule is a draft until a named reviewer signs it and a release activates it. Nothing reaches a camera unreviewed.
Corrective actions
CAPA records carrying the fields ISO 45001 clause 10.2 actually asks for, including the effectiveness review that closes them.
Verifiable reports
A report is a record, a document and a bundle: the content hash covers the record, so the same facts verify twice.
Cycle time per station
Measured from the floor cycle by cycle rather than estimated from a standard, with micro-stops surfacing as they happen.
SOP adherence
Step order, visibility and timing against a registered reference. A prototype today, validated on one demonstration workflow, not on a customer line.
Material flow & inventory
Objects as well as people: what moved, where it went, and what is actually line-side.
Availability labels: Available · Pilot (running with pilot customers) · In development. Readiness is stated per capability, not per product — a repository existing is not a release.
The evidence chain
Built so a stranger can check it.
An incident record is only worth the scrutiny it survives. Four decisions make Sight’s records checkable by someone who does not have our software and does not take our word for anything.
The record is the artefact
A report is JSON first; the document is a deterministic rendering of it and the PDF is never part of the content hash. Regenerate the same report and the hash matches, because timestamps and font subsetting live in the envelope, not the content.
Deletion leaves a stated hole
Retention purges the picture and keeps its digest: “purged under policy on this date, digest was X” is a complete answer. “File missing” is indistinguishable from tampering. Legal hold stops the clock.
Every write is attributable
Every principal is named and every write carries one. There is no fallback actor: an unauthenticated change is refused, never recorded as anonymous. The audit chain outlives the evidence it describes.
Models are admitted, not uploaded
A model profile is sealed and passes explicit admission gates before it can be assigned; an eligible candidate is still not an active one. Evidence that feeds retraining stops at a reviewed outbox.
Security & privacy
On-premise by default. And a straight answer on identity.
Processing happens at the edge inside your plant. Event streams are encrypted, video never leaves the site, and we store events rather than footage.
Sight does track people, and we would rather say so than be caught wording around it: you cannot tell whether one person crossed into a restricted area without a helmet unless you can follow that person across two cameras. What Sight keeps is a track identity — a handle that lives for as long as the rule needs evaluating. It carries no name, there is no face-recognition model anywhere in the stack, and it is not joined to your HR system. Findings name a zone, a time and a rule, and they go to the person accountable for that zone. We publish this because it is the first question a works council asks, and the answer should not be a surprise.
Video never leaves your plant
Cameras are processed on an edge box inside your network. Footage stays on site.
Events, not footage
What crosses to dashboards and audit logs is timestamps, zones and rule outcomes, not video.
Air-gapped when you need it
Cloud-managed, your own VPC, or fully offline on-prem. Your choice, your data.
A track id, never a name
No face recognition, no HR join. The identity exists to apply a safety rule once, then expires.
Why vision
Why Sight makes the rest of the platform smarter.
Software can only be as smart as its picture of the plant. The same perception stack that proves a safety rule is the one that will measure a cycle: detection, tracking and re-identification are the hard part, and the rules on top are comparatively cheap. That is why the order is deliberate — we are earning the right to the harder claims on evidence that already stands up. And because supervisors confirm or correct what Sight flags, every correction becomes a reviewed candidate for the next model release, never a silent change to the one in production.
Six weeks to live data