Your construction site cameras already record everything. Can you search it?
Most sites already pay for cameras, and the footage gets opened after something has gone wrong. I built a platform where a site manager types a question and gets the matching frames back.
By Firas Mohammed · · 3 min read
The short answer
On most sites, no. The cameras record every day, and the footage gets opened when there's a dispute, by someone scrubbing a timelapse by hand. I built ConstructVision to fix that. Every frame is stored with a semantic embedding, so a site manager can type what they're looking for and get the closest frames back. I designed and shipped it in about three weeks, directing AI coding agents. It runs on staging and has no paying customer yet.
What the search looks like
A project manager wants to know when a particular pour happened, or whether scaffolding was up in a given week. Today that means dragging a slider through weeks of frames.
In ConstructVision you type the question. Each frame is stored with a 768-dimension SigLIP2 image embedding in PostgreSQL with pgvector. A query like "concrete pour near the east crane" returns the nearest frames, scoped to one project or one camera.
Two detectors run on every frame. One is a standard object model limited to site classes. The other is an open-vocabulary model reading a short construction word list, including crane, excavator, scaffolding, dump truck and "worker wearing hard hat".
The same record feeds the things people forward: PDF reports with charts and an optional written summary from the Claude API, timelapse video exports, a web portal and an Android app.
Detections on one test frame after I raised the detector's input resolution from 640 to 1280 pixels.
Where it sits
Partway through the build I had the product torn apart against its competitors. The premium tools, OpenSpace, Buildots and Doxel, map site captures to a BIM model and a schedule, and charge from USD 10,000 a project. Fixed-camera vendors sell hardware, connectivity and timelapse, with AI added as motion and PPE alerts.
I aimed at the space between them: a fixed-camera product that answers questions in plain language, on cameras a site already owns.
What a buyer should check
If you run sites in the GCC and someone pitches you camera AI, these five checks take an afternoon.
- Cross-tenant probe. Log in as one client and request another client's frame, camera or report by its ID. The answer should be the same "not found" you get for an ID that doesn't exist. ConstructVision returns not-found here, so a probe can't learn that the other client's data exists at all.
- Viewer limits. ConstructVision has three roles (system admin, client admin and client viewer) over a hierarchy of clients, projects, cameras and frames. Give a viewer account to someone outside the project team and ask them to change something. The server should refuse every attempt.
- The camera upload path. Ask how a camera proves who it is and what happens to a file before it's stored. Mine gives each camera its own token, re-encodes every image to strip EXIF data and anything embedded in it, accepts only listed content types and caps the file size.
- Shared links. If the product shares frames with outsiders, set a link to expire, switch downloads off, then request the full-size image directly. The server should refuse.
- The release gate. Ask the vendor to run their release checks on a clean copy of the code while you watch. Mine runs backend tests on locked dependencies and proves that a fresh database and an older one migrated forward end up with byte-identical schemas. Then it typechecks, lints and builds the frontend. CI runs the same gate.
Questions people ask
Can AI search existing construction camera footage?
Yes, if every frame is stored with an image embedding. ConstructVision does this for fixed-camera frames. Its accuracy on a live site hasn't been validated yet, so treat any count it shows as a lead to check by eye.
Do I need a BIM model to search site footage?
Search works on the images alone. Measuring percent complete against BIM and a schedule is what the premium tools charge for. ConstructVision doesn't measure percent complete.
How many cameras can it handle?
A handful, today. Inference runs on CPU, one frame at a time, which is too slow for a hundred cameras. The planned fix is a job queue with a GPU worker, and the limit is written down in the project's own risk register.
Is ConstructVision available to buy?
It's deployed to staging. It went through a client-readiness audit and I fixed the findings. It isn't running for a production customer today.
The case study
The full build, including the competitor teardown that listed what the product couldn't do, is in the ConstructVision case study. I ran the same kind of outside-in read on pay-at-table apps in Kenya.
If you need something like this built for your sites, here's what I take on.