Skip to main content

Kestrel AI Early testing

UK-built AI for safety inspections.Built with UK inspectors.

Kestrel is SiteSurvey's inspection AI. It finds the parts of a fire door live through the camera, locks onto them as you move, measures them and checks each one against UK guidance. Today it runs on Google's Gemini models. Next, our own Kestrel models, trained on UK examples and tested by UK inspectors. We are looking for founding partners to build them with us.

Kestrel systemEarly testing on iPhone: live AR detection, LiDAR measurement and safety rules, on Google Gemini models.

Kestrel modelsIn development: our own models, trained on UK examples.

In the appThe AI inspector, in the SiteSurvey app now, on every account.

Screen recording: a fire door point is filmed, the AI suggests Fail with the torn intumescent strip boxed in red, and the inspector accepts it
Torn seal foundInspector confirms

What is Kestrel AI?

Kestrel is SiteSurvey's inspection AI for compliance and health and safety defects, starting with fire doors. The Kestrel system is in early testing: it detects a door's parts live through an iPhone camera, locks onto them with AR, measures them with LiDAR and checks them with code-enforced safety rules, using Google's Gemini models for recognition. Our own Kestrel models are in development. They will be trained on images shared by partners, tested against verdicts agreed by certified inspectors, and have their accuracy published before they replace Gemini for any customer.

Last updated . Facts checked by the SiteSurvey team on 28 October 2026.

Working today · early testing

It finds the door, locks on and measures it.

On an iPhone test build, Kestrel boxes the parts of a fire door live while you film, pins each box to the real door with AR so it stays put as you walk up, and measures the door with LiDAR. When you finish the point, it picks the sharpest stills and checks each item.

Finds the parts live. While you film, the door, frame, signs, closer and anything holding the door open are boxed as they come into view.

Locks onto the real door. On iPhone, boxes are anchored in the room with Apple ARKit, so they grow as you walk closer and stay put when you pan away and back.

Measures it. LiDAR gives each item a real size, and the size stays locked once found.

Reads signs properly. A sign only counts as "Fire door keep shut" when those words can be read. Any other sign is marked as another sign.

Never guesses a gap. A gap passes or fails only from a legible gap gauge or caliper reading. Everything else is "measure on site".

The clip is AI-generated for testing, not a real inspection. The live boxes are real Kestrel detection, about twice a second. Recognition by Google Gemini; detection, tracking, measurement and rules built by SiteSurvey.

How partners help

Three things we need. Four things you get.

1

Defect categories

Help describe and grade defects the way your guidance and members do, so Kestrel speaks the same language.

2

Images

Example photos or video of real defects, shared under a data-sharing agreement and anonymised before use.

3

Expert reviewers

Certified inspectors to check a sample of Kestrel's verdicts and tell us where it is wrong.

✓

What partners get

  • Early access for your members or teams
  • Published accuracy, area by area
  • Anonymised defect trends
  • Credit as a founding partner

First results, October 2026

What the approach caught on real doors and phones.

Early tests of today's AI inspector following our UK briefs. Not a formal trial and no accuracy figures yet: they show the approach works. Proving how well it works is what we need partners for.

AI inspector findings: fail, the intumescent strip on the hinge jamb is torn, boxed on a still
Fail

Torn intumescent strip

On the seals point of a fire door, boxed on a still from the clip, with a request for a fuller shot of the hinges.

AI inspector findings: fail, fire door held open with a wooden wedge
Fail

Door wedged open

On the closer point. It asked the inspector to remove the wedge and film the door closing from fully open.

Stills saved with the finding, the defect boxed in red
Evidence

Stills saved with every finding

The frames behind each verdict are kept with the form, so the finding can be checked and defended later.

Fail

Fire door on a real siteThe door did not self-close when released.

Fail

Electrical visual checkA light switch with a missing rocker, marked critical.

Fail

Quick defect checkA charger unit missing from its mounting, boxed on the still.

Cannot assess

Points not filmed fullyIt said what was missing (the gauge, the reverse face, the seals with the door open) instead of guessing.

Why a UK model

General AI was not trained on UK buildings.

Today's vision models learned mostly from overseas images. They have seen few UK fire door sets, intumescent seals, consumer units or HHSRS hazards, and they do not know UK guidance.

Inspectors need consistent evidence they can defend. Regulators need to know how a finding was reached. Kestrel aims to do both: trained on UK examples, labelled against UK guidance, tested by UK inspectors, with its accuracy published.

How it works · Now live

Film the point. The AI checks it. The inspector decides.

A guided Walk Cam takes the inspector through each point of an inspection. Each clip is checked against a written brief for that point, and seconds later the inspector sees a suggested pass, fail or cannot assess with any defect boxed. Kestrel is designed to slot into the same flow.

Screen recording: the inspector films point 1 of 4 on a fire door

1. FilmEach point the app asks for.

Screen recording: the findings sheet shows Fail with a torn intumescent strip boxed in red

2. CheckA suggested verdict, with the defect boxed.

Screen recording: a door wedged open is flagged and the inspector taps Accept Fail

3. DecideThe inspector accepts, changes or dismisses.

Screen recording: the form shows each point with its confirmed verdict and boxed photos

4. ReportOnly confirmed findings reach the report.

What Kestrel will show

The spotter finds it. The inspector model judges it.

Kestrel cuts each clip into stills and keeps the best one of every item. The spotter boxes what it finds in light blue. The inspector model then colours each item by its verdict and writes what to do about it.

Three versions of the same fire door: the photo, the spotter's light blue boxes with confidence scores, and the inspector's colour-coded numbered marks

Illustration on our own fire door footage, drawn to show the layout of results. It is not output from a model.

Every finding comes with an action. Fail, check or pass, the reason, and what to do next: fit a sign, measure the gap with a gauge, confirm the spyhole is fire rated.

It asks when it cannot see. The hinges were not in this shot, so it asks for a close-up of the hinge side instead of guessing.

Readable for everyone. Each verdict has its own symbol as well as its colour, so it works for people with colour blindness and on a black and white printout.

See every step, from clip to verdict →

Follow one fire door through all four points, to the finished report →

FailA defect that needs fixing. Says what to do.
CheckSomething to confirm or measure on site.
PassSeen and fine.
Better shot neededNot visible enough to judge. Says what to film.
Spotter boxLight blue, with a name and how sure it is. Found, not judged.
ObjectA box round a closer, hinge, sign or handle.
Gap or sealA traced line along long, thin things.
MissingDashed where something should be but is not.

Everything we want Kestrel to cover

Fire doors first. Then every kind of inspection.

Each area gets its own brief: the checks for each point, the guidance it follows, and the rules for pass, fail and cannot assess. Kestrel checks what can be seen. It never replaces testing, sampling or a qualified person's judgement.

LiveRuns today on every account, from a library form.TestedRun on real phones, not in the library yet.PlannedNeeds partners, images and reviewers.

Fire safety

Fire doors are the first worked example: 4 points and 18 checks covering door and signage, gaps, seals and hinges, and the closer. Communal fire door checks have their own brief with 17 checks.

  • Fire doorsLive
  • Communal fire door checksLive
  • Escape routes and fire exitsPlanned
  • Fire stopping and compartmentationPlanned
  • Emergency lightingPlanned
  • Fire extinguishersPlanned
  • Fire alarm call points and detectorsPlanned
  • Fire safety signagePlanned
The AI inspector suggesting a fail on a fire door seal, with the defect boxed

The AI inspector on a real fire door.

Missing an inspection your organisation cares about? Tell us and we will add it.

How we will prove it works

Tested by inspectors, published in the open.

An evaluation set checked by certified inspectors

Real UK inspection clips, each verdict agreed by qualified people before Kestrel is scored against it. Kept separate from anything it is trained on.

Wrong passes reported on their own

A missed defect is the failure that matters. Wrong passes are counted and published separately from overall accuracy.

Results published for every area

Accuracy for fire doors, ladders, electrical and each new area, published in the open, including where it falls short.

No customer use until it passes

Kestrel stays out of the app for an area until it meets the agreed bar on that evaluation set.

A model that misses a defect is worse than one that asks for a better shot.

The inspector decides

Every finding is accepted, changed or dismissed by a competent person. Nothing reaches a report until they do.

It says when it cannot see

Out of shot, too dark, or no gauge in the gap all come back as cannot assess, with the shot that would settle it.

No measuring by eye

It reads a gauge or ruler on camera. It never estimates a gap or a distance.

Data handled properly

A data protection impact assessment, written data-sharing agreements, anonymisation before training, and nothing used without permission.

The plan

From today's inspector to a UK model.

  1. Now · we are here

    Kestrel system, early testing

    Our own flow, live AR detection, LiDAR measurement and safety rules, running on Google Gemini models for recognition. The AI inspector is in the SiteSurvey app now, on every account. The inspector confirms every finding. Clips are not used for training.

  2. Next

    UK Safety Image Library

    A shared, anonymised set of UK inspection images, labelled with defect categories agreed with regulators and trade bodies, contributed by partners under data-sharing agreements.

  3. Then

    Kestrel models, UK-trained

    Our own models trained on that library, starting with fire doors, tested against the evaluation set before they replace Gemini for any customer.

  4. Always

    Results in the open

    Published accuracy for every area, with wrong passes reported separately.

The detail, including the design, base models, data and testing: read the Kestrel technical plan →

Help us build it

We are looking for founding partners.

Regulators, trade bodies, fire and rescue services, housing providers, contractors and universities. Tell us a little about your organisation and we will start with a call.

Regulators and trade bodies

Shape the defect categories and grading so Kestrel uses the language your guidance and members use.

Housing providers, contractors, fire services

Share example images under agreement and pilot it on real inspections.

Certified inspectors and universities

Review a sample of Kestrel's verdicts and help design how it is tested.

What partners get

Early access for your members or teams. Published accuracy, area by area. Anonymised defect trends. Credit as a founding partner.

Talk to us about partnering

A few details and Lee Moss, our founder, will reply personally. No commitment: most partnerships start with a call.

How could you help? (optional)

Prefer email? lee@sitesurvey.app

Frequently asked questions

Kestrel AI: common questions.

Is Kestrel available yet?

In early testing. The Kestrel system runs today on a test build and demo accounts: it finds a fire door's parts live through the iPhone camera, locks onto them with AR, measures them with LiDAR, and checks each one with code-enforced safety rules. Recognition currently uses Google's Gemini models. Separately, the AI inspector is in the SiteSurvey app now, on every account. Our own Kestrel models, trained on UK examples, are in development and will replace Gemini area by area once they pass testing.

Is Kestrel your own AI model?

Kestrel is SiteSurvey's own inspection AI system: the flow from clip to verdict, the live AR detection and measurement on the phone, and the safety rules are built by us. The model that recognises what is in each picture is Google's Gemini today. We are building our own Kestrel models, trained on UK inspection examples, to replace it. We will say so clearly when each one does.

What will Kestrel inspect?

Fire doors first, then ladders, electrical fittings, communal areas and escape routes, damp and mould, and site safety. The full list, with where each one stands, is on this page.

Will my clips be used to train it?

Not unless you choose to contribute them. Clips checked by the AI inspector today are not used to train any model. Kestrel will be trained only on images partners share under a data-sharing agreement or that customers explicitly opt in to share, anonymised first.

How are partner images handled?

Under a written data-sharing agreement with each partner that sets out how the images are anonymised, stored and used. Faces, addresses, flat numbers and number plates are removed before anything is used, and a data protection impact assessment covers the work.

How will you know it is accurate?

Kestrel will be scored against an evaluation set of UK inspection clips whose verdicts are agreed by certified inspectors and kept apart from the training images. Results will be published for every area, with wrong passes reported separately, and it will not be used by customers in an area until it meets the agreed bar.

Does it replace a qualified inspector or a certificate?

No. It suggests findings to a competent person, who accepts, changes or dismisses each one. It checks what can be seen on camera. It does not replace testing, sampling or a qualified person's judgement, and reports say "AI-assisted inspection. Verdicts confirmed by the inspector."

How can my organisation get involved?

Regulators, trade bodies, fire and rescue services, housing providers, contractors and universities can help with defect categories, example images, expert review or pilots. Use the form on this page or email Lee Moss, our founder, and we will start with a call.

About SiteSurvey

SiteSurvey is a UK inspection and survey app used by housing providers, contractors and facilities teams. Inspectors record forms, photos, Walk Cam video and floor plans on their phones, and the office gets the report. Updated 28 October 2026.