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.
AI Consulting
Which step of a workflow genuinely needs a model, which one computer vision settles, and which decision stays with a person? That split is made before the work reaches a model, not afterwards.
AI consulting means deciding which step of a job needs a model and which one goes to computer vision. The split comes before any model call: the threshold for a filled mark sits in the code, and a model enters only where meaning or matching is needed. Vipkurum's optical marking tool runs no language model at all: marks are read from the form image, and a branch scores its own exam.
Optical marking used to be a hardware job. The reader sat at head office, branches carried their sheets to it, and on the day it broke marking waited. Nor could anyone step outside the form the device would accept.
At Vipkurum that dependency was removed by design. The reading happens on the image of the form rather than in the optics of a device: whether a field counts as marked is decided by the software’s image analysis layer, and we go through that build step by step in scoring exam sheets without an OMR device. When the form layout changes, the adjustment is made in software, so there is no new device and no device-specific form to order. A branch marks its own exam.
Not one line of a language model runs in that job. The model enters the next one: when a question booklet is cropped, computer vision sets the boundaries while the language model matches the answer key and writes new questions in the same shape. The two jobs are kept apart on purpose.
Where a model does run, the last word still belongs to a person. At Tahlil.com AI prepares the first reading and a specialist physician writes and approves the interpretation, so the text that reaches a patient is signed by the doctor who approved it. Which decisions a model will not make is written down before the model runs.
This page is not offering an AI strategy. There is something to talk about when a concrete step, a pile of existing material or a measurable output is on the table: a mark to be read off a form, a question to be lifted out of a booklet, a piece of text that has to go for approval. If none of your steps look like that, we will say so plainly.
The AI work divides by which step genuinely needs a model:
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.
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.
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.
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.
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.
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.
Every project goes through the same four stages: planning, architecture, modelling, development.
See our process
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.
Marks on an optical answer sheet are read from the image of the form rather than through a reader's optics. Scoring stops being something that happens at one machine in one office: whoever uploads the sheet gets the result.
Our own product, still in development. Someone creates an AI character, and that character posts into a shared feed. Its identity and the history it builds up sit on an account rather than inside a chat session, and text and images come from two separate generation paths.
OCR reads the letter shapes in an image and returns text. Which reading a job needs follows from the document: letters, a mark with a fixed place, or a region to separate.
The same question comes in daily; whoever reads the messages that day answers it. Whether a bot can take it depends on where the answer lives, not on its topic. No rule, and the bot improvises.
An OMR device sits in one office; branches ship their answer sheets to it and wait. Move the reading from the device's optics to the form image, and every branch that can scan can score.
A model produces output; someone still answers for it. From lab-result interpretation to exam scoring, our projects show where the approval step sits and why oversight is a founding decision.
The wait a user sees comes from the request the job never left. One flow step from Secretum, which we are building now: what gets queued, where the file lands, how the result returns.
It does not, and the first thing we do is find out which step actually needs a model. That keeps you from paying for model calls on a job that image analysis settles on its own. Vipkurum's optical marking tool runs no language model at all: marks are read from the form image, and the decision threshold sits in the code. Walk us through your process step by step and we will separate the model work from the vision work with you.
Your booklets turn into a question bank where every item carries its own answer. Your teachers stop cropping in an image editor and typing keys by hand. Vipkurum's question cropping tool splits a booklet page into individual questions with computer vision, and the answer key is written onto each cropped question's own record; one slipped row no longer turns a whole exam's marking wrong. The questions you already hold become the pattern for new ones; send us a sample booklet and we will show you the result.
We can, and it does not mean rewriting what you already run. The vision or model work lives in a separate service; your existing application calls it and writes the result into its own record, so your screens and your database stay where they are. Vipkurum's optical marking runs as a separate reading service, and its results land in the exam record the platform already keeps. Describe your system to us and we will come back within 2 business days.
We manage it in two places: a named person owns the output, and the model's limits are drawn before it runs. Whatever reaches your users has someone accountable behind it, not a model's unchecked interpretation. Tahlil.com has AI prepare the first reading, and a specialist physician writes and approves the interpretation before it reaches the patient. Vipkurum's question cropping tool leaves the question boundaries to computer vision, never to a model. We settle which decisions stay with a person in the first conversation with you.