What it compares to
You put someone in front of the material to watch it and take notes.
That is not a bad method. A person understands at once what they see, spots connections no machine spots, and decides the borderline cases. You can also tell them what to look out for. In that sense a person is more flexible than any software. For a single recording, it is the fastest way there is.
It just does not scale. And it leaves no trace.
| Someone watches it | The application | |
|---|---|---|
| Effort per recording | About the length of the video, usually more, because of rewinding. | Runs in the background, with nobody sitting by. |
| With twenty recordings | Twenty times the effort. Concentration drops, short appearances get lost. | Frame 1,800 gets the same attention as frame one. |
| Looking for something else | Watch everything again, with a new task in mind. | Change the question in the text field, run it again. |
| The result | An assessment. | Evidence image and timestamp per finding. |
| Further processing | Typing it up. | Figures sit in a database and can be calculated further. |
| Where the data sits | On the person's computer. | In a database in Europe, under your own access. |
One hour of footage at two seconds apart is 1,800 single frames. The application asks the same question 1,800 times. As precisely at the last frame as at the first.
What concretely runs today
Someone uploads a video file and sets how finely it should be checked.
The application splits the video and sends every single frame, together with the stored question, to the AI model. What comes back is not a yes or no, but a description: what is visible, how clearly, in which situation, with what confidence.
Findings from different frames are merged. The same object from two camera angles counts as one, not two. For this the AI compares the images with each other, not their descriptions.
For every object it calculates how often it appeared and for how long in total.
One click creates a PDF: a card per object with evidence image, duration, appearances and every find with its timestamp.
On top sits an assessment layer: you store your own documents as a rulebook, the application derives which findings are relevant and judges each one against it, with reasoning, counterarguments and a rating.
Point 6 is the decisive one. It shows that the report is not the end of the chain but a waypoint.
What you can connect
Who we built this for and what they search for is not stated here. That is not a hole in the text but the proof of the point: the question is a setting, not a component.
After the report, the same pattern always follows. The objects found, with their duration, meet something stored: a rate card, a contract, a list, a rulebook. Out of that comes a number or a verdict. This very layer is already built in the running project, for a rulebook made of the client's own documents. For a different use case, what changes is the reference and the formula, not the application:
What could be connected
| Area | Examples |
|---|---|
| File storage | SharePoint, Google Drive or a network drive |
| Mailbox as input | File in by email, report back by email |
| Calendar | Tie evaluations to appointments or campaign periods |
| Inventory and CRM | Assign findings to a customer, contract or product |
Accounting and ticket systems can be connected the same way.
And only this: file storage for videos, images and reports. User management. One AI provider. A system for long evaluations in the background. PDF generation. The assessment layer described above.
- Advertising value per brand. A rate card is stored: price per second, tiered by prominence and placement. The calculation is visible time times factor times rate, summed per brand. The result is an invoice line instead of an estimate.
- Sponsorship against contract. The contract supplies the commitment, say 120 seconds of presence per event. The evaluation supplies the actual. If it falls short, proof of the difference is created. Automatically, with evidence.
- Share of attention. Not one recording but hundreds over a period. Which brand was present for how long, in which setting, with which trend. The result is a dashboard instead of a single report.
- Alert instead of report. Store a block list: content that must not be broadcast. Every hit triggers at once, with a timestamp, instead of landing in a PDF.
- Open up the archive. All findings go into an index. That makes existing footage searchable by image content instead of file name.
On top comes what works the same for each of these cases: recurring runs on a watched input folder, threshold alerts, table export, and an interface another system uses to fetch the results directly.
What it costs
By far the largest item is the AI query per single frame. Measured, that is around 1,870 tokens per image. Almost all of it goes to the image itself, the text around it barely counts. One hour of footage at a two second check interval is 1,800 images and costs under ten euros in compute. At three seconds it is a third less, at every second it is double. (As of August 2026, common vision model; tokens measured over four runs.)
Hosting and database are fixed monthly costs and do not grow with the single video. No license per seat, no minimum purchase: if you evaluate nothing for a month, you pay almost nothing that month. The cost model follows usage, not the calendar.
How it is built
A web application where upload, evaluation and reporting come together. The evaluation does not happen in the browser but in the background: split, look, group. The steps trigger each other. The report is separate and started by hand, because it fixes the result. Data, files and sign-in sit in a managed database in a European data center, under the client's own access. All AI calls go through a single place in the program. Switching the model or provider affects that place, not the evaluation logic. And the questions to the AI are not in the program code but in the database, editable through a screen of their own.
Why it runs reliably
The duration is calculated, not estimated.
No AI is involved at this point. One rule applies: if the gap between two frames of the same object is more than twice the check interval, the object was gone in between, so that is two appearances instead of one. The duration comes from the timestamps. Same video, same result, every time. And the report lists every timestamp, so you can check the math.
That sounds minor, but it decides what the result says: three times ten seconds is something different from once thirty.
The application asks instead of guessing.
If you store your own rulebook, you upload documents. The AI reads them and does not produce a finished truth from them, but two things: a summary, and a list of the questions it cannot answer from the documents. Only when those are answered or knowingly skipped does the final version get created.
Older evaluations stay untouched in the process: each one remembers which version of the rulebook it ran with. If you change the rulebook today, you do not change retroactively what last month's report rests on.
Where AI works, and where not
AI works in three places: it looks at a single frame and describes what is visible. It compares images to recognize the same objects. And it reads rulebook documents and names what is missing in them.
Deliberately without AI: splitting the video. Calculating appearances and duration. The preselection before an assessment: anything below the set thresholds is sorted out without asking an AI at all. And creating the report.
This is not thrift but a division of labor. The AI describes what is visible. The program calculates what follows from it. And everything that is calculated can be checked.
What this means for you
The reflex with a task like this is to look for a ready-made solution for exactly your own problem. And if there is none, the task counts as not doable. This case shows the other way: what got built is not a detector for one particular thing, but a chain. Split videos into pieces you can look at, question them, condense the answers into a result with evidence. What is searched for is a text field in it. What happens with the result is a calculation rule. The client in this project searches for something very specific. The application knows nothing about that. It carries out what is stored.
So the question for you is not whether software exists for your case. It is which question you would put to your material if someone watched it for you. 1,800 times an hour, without letting up.
What is searched for is not in the program. It is in a text field.
