The pieces and their boundaries are settled before code.
Several institutions, one system, records kept apart.
One codebase, both stores: iOS and Android.
We take a running system over and improve it.
AI only where the problem actually calls for it.
A preparation tool that turns a printed question booklet into individual questions. Computer vision splits the page, a language model ties each question to its answer key, and the same model drafts fresh questions from the samples already in the pool.
An education provider has shelves of printed question booklets and needs single questions out of them. Done by hand, that means cutting every question out one at a time and keeping its answer somewhere else. The tool splits the work in two, and the line it draws is not arbitrary: where a question ends on the page is something you can measure, while which answer belongs to it is something you have to read. Vision takes the first job, a language model takes the second and drafts fresh questions from the samples it already has. The tool was built next to Vipkurum’s content management layer.
We go through how the work runs today and what you expect from the software; the scope comes out of that.
We decide what the system is made of and where its information is kept.
We draw the screens and the flow before any code, and start building once you have approved them.
We write and hand over the software so another team could carry it on.
OpenCV splits a booklet page into individual questions; the boundaries come from image processing, not from a model's reading of the page.
Gemini and OpenAI models tie each key entry to its question, so the key stops being a separate document and lives on the cropped record.
Existing questions act as samples; the model drafts new ones in the same shape and produces the matching answer key alongside each.
Cropped images and question records sit in MongoDB; the Python vision service is reached through Node.js from the Laravel content management side.
| Challenge | Result |
|---|---|
| An institution owns printed booklets but needs individual questions to assemble a new test | A booklet off the shelf becomes a pool of individually cropped questions |
| Teachers crop questions in an image editor and type the answer key by hand | Cropping and key matching happen in one pass instead of two manual jobs |
| One row slipping in the key turns the scoring of a whole exam wrong | Each answer sits on its own question, so there is no key row left to slip |
| When the pool has no question for a topic, a teacher has to write one from scratch | A missing question is drafted from existing samples, with its answer alongside |
| Leaving crop boundaries to a language model makes question edges unreliable | Cropping stays with vision, interpretation stays with the model, the two do not mix |
Our own workshop software, built around the car on the ramp. A work order, the parts it consumes and the invoice it produces run as one chain, and the accounting side hangs off that same record. Maintenance history lives on the plate and chassis, so it travels with the car.
Our own product for HVAC, boiler and appliance service companies. The customer and device records the office works from are the same ones the technician carries into the field. A job can be opened and closed in a basement with no signal, and service history stays tied to the device.
Our own product, still in development and not yet live. How long an appointment runs, what it is booked against and how bookings stack all differ by trade, so we are keeping those rules configurable inside the product instead of fixing them in code.
Tell us how the work runs today and we will look at which step could move into software.
