Kestrel AI Technical plan
How we will build Kestrel.Fire doors first.
The design, the open models it is built on, the data it needs, how it will be tested and the steps to go live. Written for regulators, trade bodies, housing providers, inspectors and researchers considering working with us. Updated 27 October 2026.
In one minute
The plan on a page.
What we are building
Kestrel, a vision model that checks inspection video for visible defects. Fire doors first, then ladders, electrical, communal areas, damp and mould and site safety.
Who owns it
SiteSurvey. Kestrel is built on free, open base models whose licences let us train them, keep the result and run it on our own UK servers.
What it needs
Labelled UK examples. Around 2,000 to 5,000 inspected fire doors for a first version, plus 300 to 500 doors judged by certified inspectors to test it.
How we prove it
It runs silently beside today's AI inspector, is tested against inspector verdicts, and goes live only if it rarely passes a door that should fail.
The design
Two models working together: a spotter and an inspector.
Splitting the job in two means each part can be tested and improved on its own. The spotter can be measured defect by defect, and the inspector turns its detections into a verdict a person can read and check.
Part 1: the spotter
A small, fast detection model. Its only job is to find known problems in each frame and draw a box round them: a wedge, a torn seal, a missing closer, a missing sign, a gauge in the gap. Because it does one thing, its accuracy is easy to measure defect by defect. Small versions could later run on the phone itself.
Part 2: the inspector
A vision and language model. It takes what the spotter found, the clip and the checklist for that point, and writes the result: pass, fail or cannot assess, a short summary, and what to film if it needs a better shot. It is also what says cannot assess when the clip does not show enough.
Film a point. The spotter boxes what it finds. The inspector writes the verdict. A competent person confirms it.
Today: AI inspector
A general-purpose model (Google Gemini) follows our written UK briefs. It is the same model for everyone, it cannot learn from inspections, it runs on Google's servers, and clips are not used for training.
Kestrel
Trained on UK fire door examples labelled against UK guidance. It improves as inspectors confirm findings, it is owned by SiteSurvey, and it runs on our own UK-hosted servers. Nothing changes for the inspector: same walk, same buttons.
Built on open models
We do not build from scratch. We train free open models and own the result.
Training a vision model from nothing takes hundreds of millions of images. Instead, Kestrel starts from free models that already understand everyday images, then learns UK fire doors from our labelled examples. This is called fine-tuning.
Inspector base: Mistral Small
Made by Mistral AI in Paris. A 24 billion parameter model that reads images and writes text. Licence: Apache 2.0, which allows commercial use, training and keeping the trained model private, with no fee or royalty. It needs one large cloud graphics card to run.
Spotter base: RF-DETR
Made by Roboflow in the US. A real-time object detector. We will use only the Nano to Large sizes, which are Apache 2.0. The XL and smallest sizes carry a different licence and are not used. It runs on a basic graphics card.
Who owns Kestrel
SiteSurvey. Mistral and Roboflow keep their original free models, which anyone can download. The trained Kestrel model, our data and our labels are ours. We keep their licence and credit notice with the model, as the licence asks.
Why these two
Licences that let us own and host the result in the UK, strong performance on images, and makers based in Europe and the US. Base models improve quickly, so we will re-check before each training run and only switch to one with an equally clear licence.
Scope of the first model
The fire door defects Kestrel will look for first.
Today's fire door brief has four filmed points and 18 checks: door and signage, gaps, seals and hinges, and the closer. The first version targets the defects below. The final list and grading will be agreed with fire door trade bodies.
Door held or wedged open
Wedges, hooks and objects stopping the door closing.
Door does not self-close
The door is released and does not close fully into its frame.
Damaged or missing seals
Intumescent strips and smoke seals torn, painted over, loose or missing.
Gaps too large
Read from a gauge or ruler held in the gap on camera. Never estimated by eye.
Signage missing or damaged
Fire door keep shut and similar signs on the correct face.
Leaf or frame damage
Holes, splits, delamination and damage to the frame or architrave.
Hinges and fixings
Missing screws, damaged or missing hinges, and visibly loose fixings.
Closer and hardware
Closer missing, detached or visibly damaged; damaged handles and latches.
What it will not judge
Fire rating and certification
A camera cannot tell whether a door is a certified fire door. Kestrel can read a label or plug if one is filmed, and says cannot assess if not.
Measurements without a gauge
It reads a gauge or ruler on camera. It never estimates a gap, a length or a distance.
Anything hidden
Concealed hinges, the inside of the frame, or anything not in shot. These come back as cannot assess, with the shot that would settle it.
The data
Thousands of labelled doors, not millions of images.
Training the model is the easy part. Collecting and labelling enough real UK fire door footage is the real work, and it is where partners make the difference. These figures are planning estimates.
Test set: 300 to 500 doors
Each door judged separately by two certified fire door inspectors. Never used for training. This is the test Kestrel has to pass.
First version: 2,000 to 5,000 doors
Enough for around 500 good examples of each of the defects above. One inspected door is about four clips, dozens of frames, and a verdict on every check.
Full coverage: 10,000 doors and more
Brings in the rare defects, which may appear only once in hundreds of doors.
Rare defects on purpose
Scrap or test fire doors filmed with deliberate faults, alongside real footage, so the rare cases are covered. Generated footage is never used to test the model.
Where it comes from
SiteSurvey customers who opt in
Every accept, change or dismiss on an AI finding is an example labelled by a real inspector. Used only with the customer's explicit permission.
Housing providers' photo archives
Inspection photos with defect notes from past reports, shared under a data-sharing agreement and relabelled.
Certified inspectors
Paid to label examples and to judge the test set.
Trade bodies, test facilities and manufacturers
Defect categories, reference images and staged defects from test doors.
A label is more than a note. Each example records a verdict for every check and a box round each defect, so the model learns exactly what and where. Free-text notes from past reports are useful, but are relabelled before training.
Testing
It goes live only when it rarely misses a defect.
Wrong passes come first
A false pass, saying a door is fine when it is not, is the number that matters most. Our target for critical defects is under 2%, and never worse than today's model.
Shadow mode before live
Kestrel runs silently on real inspections first. Inspectors see only today's results until Kestrel has proved itself.
Two inspectors, not one
The test set is judged by two certified inspectors separately, so the benchmark itself is reliable.
Results in the open
Published accuracy for each defect, with wrong passes and wrong fails reported separately.
The phases
From foundations to a live fire door model in about a year.
- Months 0 to 2 · next
Foundations
Agree the defect list and grading with trade bodies. Build the consent, anonymisation and labelling pipeline. Start the test set.
- Months 2 to 6
First version, in shadow
Train the spotter on the main defects and the inspector on verdicts. Run Kestrel silently beside today's AI inspector on real walks and compare both with what inspectors confirm.
- Months 6 to 12
Fire doors go live
Switch fire doors to Kestrel only if it passes the test set and beats today's model. Publish the results.
- Months 12 to 18
All checks, next area
Cover every fire door check, then use the same pipeline for the next area, starting with ladders or communal areas.
Timings depend on partners and data. Each phase starts once the one before it is complete.
Data protection
Handled properly from the start.
Permission first
Nothing is used for training without a written data-sharing agreement or a customer's explicit opt-in.
Anonymised before training
Faces, flat and house numbers, addresses and number plates are blurred or removed.
A data protection impact assessment
Completed before any partner data is collected, and shared with partners.
UK hosting
Training data and the Kestrel model will be kept on UK-hosted servers.
Help us build it
Defect categories, images, expert reviewers and pilots.
If your organisation can help with any of these, we would like to talk.