Get a quote

AI Consulting · Kadikoy

AI Consulting for Kadikoy businesses

Cropping by hand fits into the gap between two appointments, but it never comes out the same twice. The margin drifts, one cut takes the top line of the item below it, and next week somebody else picks the job up and the measurement moves again. If you break printed material up that way in a clinic, a studio or a small shop around Kalamis and Sahrayicedit, the measurement drifts a little each round. In the tool we wrote for Vipkurum, image processing draws the cut instead of a hand.

AI Consulting — Kadikoy

Kadikoy leans on clinics, studios and small retail that runs by appointment. If your work is organised around appointment slots, splitting a page into pieces falls into whatever gap is left between two customers, and that gap is never the same length twice. The work closest to this district's sector profile is Vipkurum's question cropping tool, built in an education technology context. Where it does not overlap belongs on the page: a booking diary, a patient file and a treatment plan are outside that tool, and we promise nothing where we can show nothing. Python and OpenCV write the layer that draws the boundary, Gemini and OpenAI models handle the matching, cropped images sit in MongoDB, and the service reaches Laravel through Node.js.

  • Image processing makes the call on the cut: OpenCV reads each question's edge out of the geometry of the page itself.
  • Feed the same page again and it separates along the same lines, so the measurement does not move with whoever picked the job up.
  • The scope is narrow and written down: a booking diary, a patient file and a treatment plan stay outside this tool.

Why Algoriel

  • The image does the reading

    Vipkurum reads optical form marks with OpenCV from the image of the sheet, not through the optics of a reader appliance.

  • Cropping apart from judgement

    Image processing decides where a question begins and ends; the language model only binds the answer to it. The two jobs never mix.

  • Model output lands in a record

    Vipkurum writes the answer key into the cropped question's own record instead of keeping it as a separate document.

  • A person has the last word

    At Tahlil.com the model drafts the reading and a doctor approves what reaches the patient; an unapproved reading never appears on screen.

What we build

AI Consulting
  • Where no model is needed

    Marks on an answer sheet are read from the form image itself; the threshold that decides whether a bubble counts as filled sits in the code, and no language model runs in that step.

  • Splitting the job in two

    Computer vision sets the question boundaries; a language model then ties each answer in the key to its question. Cropping accuracy never rests on a model's interpretation.

  • Removing the hardware step

    Marking stops depending on a reader device; when the form layout changes, we adjust the reading service in software instead of ordering new hardware or device-specific forms.

  • Turning existing material into a pool

    A booklet page is split into individual questions, the answer key is written onto each question's own record, and the questions already on hand become the pattern for new ones.

  • Tying the output to a person

    At Tahlil.com AI prepares the first reading, a specialist physician writes and approves the interpretation, and responsibility for the text that reaches the patient stays with the physician who signed it.

  • Queueing the generation

    In Secretum text and images come out of two separate generation paths; image generation is not instant, so that work goes onto a queue, and the finished file is stored once instead of being regenerated in the feed. Secretum is the product we are building now.

Work like this

AI Consulting · Case study

Vipkurum question cropping — the page's geometry sets the edge, not a hand

The question cropping tool we wrote under the Vipkurum roof. As a booklet page is separated into single questions, the place of each cut is set by the geometry of the page rather than by a hand: Python and OpenCV read the edges out of the image. Upload the same page again and it divides along the same lines, so the margin does not move with whoever picked the job up. Cropped images and question records sit in MongoDB, and the image processing service reaches the Laravel content side through Node.js. The scope is education material; a booking diary, a patient file or a treatment plan are not work we can show.

Read the case study →

Working with Kadikoy teams, remotely

We have no office, so work for this city runs remotely and we do not offer on-site support. What was done, and why it was done that way, stays in writing. The project pages mark which builds came out of a partnership.

  • The first release ships as a working piece; extras follow later.
  • Every client we know of is in Turkey.
  • Decisions stay in writing; progress shows in working screens.

Frequently asked

  • Our material is a printed catalogue page, not a question booklet. What does this say to us?

    The tool was built for education material and we do not hide that. What carries across is not the subject but where the cut comes from: edge lines, white space and block heights. For catalogue management or a product card we have nothing to show.

  • Is a model doing the splitting?

    Image processing does the splitting. The boundary comes out of the page's own geometry, so the measurement does not depend on who started the job. The model works somewhere else, not on this step.

  • Our pages scan crooked and the edges are not clean. Is that a blocker?

    Because the measurement is read out of the page's own geometry, what matters is where the edge actually falls on the paper, not a fixed template. We put that forward as a step to test together rather than a promise; we hold no measured success figure.

  • Do we have to send you the material?

    We keep no office and the work runs entirely remotely. Nothing physical travels to us. We work over the image of the page and write down together which block counts as a record of its own.

Let us look at which step needs a model and which one needs computer vision.