Our work / Vision AI
A system that reviews camera footage of a procedure, checks every step against the written SOP, and gives the quality team a timestamped record of anything missed or done out of order.
What the model caught
- Hand wash or sanitiseNot observedoptional
Fig. 1 Replay of a real gowning review (3:13). It starts when it scrolls into view and logs each step as it happens. Select any step to jump to that moment.
- Input
- Live CCTV streams or recorded video
- What it checks
- Whether each step happened, in the right order, done the right way
- Output
- A list of deviations, each with its timestamp and clip
- Where it runs
- On your own servers or a private cloud
The problem
Many operations depend on people following a written procedure, step by step and in a fixed order. A cleanroom gowning routine, a line clearance, a safety check before entering a zone. When a step is skipped, the cost can be a contaminated batch, an audit finding or an injury.
Most teams check this in two ways. A supervisor watches for a few minutes on a shift, or someone reviews footage after something has already gone wrong. The cameras record every minute, but nobody has time to watch them, so most deviations are never seen.
Why it needs more than object detection
Standard camera analytics can tell you that a person is wearing a mask in a single frame. An SOP asks harder questions.
The order matters.
A frame can show a coverall being worn. It cannot show whether the gloves went on before the mirror check. The system has to follow the whole sequence.
Steps look alike.
A coverall half zipped and fully zipped can differ by a few centimetres of fabric.
Real footage is messy.
Several people share a room, cameras sit at awkward angles, and people step out of frame and come back.
How we built it
Turn the SOP into checks
We break the written procedure into its steps, the order they must follow, and what a correctly done step looks like.
Follow each person
The system keeps track of each person through the clip, so every step is credited to the right individual.
Check every step over time
A vision-language model describes what each person is doing across short windows of video. A rules layer compares that with the SOP and marks each step as done, missed, out of order or done wrong.
Show the evidence
The results page plays the video with every flag on a timeline. A reviewer can confirm or dismiss each one in seconds.
Under the hood
- An open vision-language model, served on a single GPU server, so footage never has to leave the site
- Works with standard RTSP and ONVIF cameras, or with recorded files
- Frame sampling tuned to how quickly each step happens, to keep processing costs low
- Every result and every reviewer decision is stored, giving the team an audit trail
Choices we made on purpose
A person makes the final call.
The system finds possible deviations. Someone from the quality team confirms them. That keeps the results trustworthy and the team in control.
Accuracy is measured on your footage.
Camera angles, lighting and uniforms differ from site to site, so a generic accuracy figure would mean little. We tune the checks on your own footage and review the results with your team before anything goes live.
New SOPs are configured.
Adding a procedure means describing its steps. It does not mean collecting thousands of labelled images and training a new model.
The first SOP: cleanroom gowning
We started with gowning because it is one of the hardest procedures to check by eye. Several garments go on in a strict order, the differences between steps are small, and a mistake can contaminate a batch.
On a typical clip, the system flags issues like these:
- Self-check at the mirror before the gloves went on, flagged at the exact second it happened
- A step that never happens, such as hand washing, shown as not observed
The same approach applies to any procedure that people carry out in view of a camera.
- Hand hygiene
- PPE before entering a zone
- Line clearance
- Assembly sequences
- Food handling
- Equipment start-up checks
Have a procedure that needs checking?
Describe the process or send us a short clip. One of the founders will reply within a working day.