- Case 06 · EdTech
We use our own product to show how we build larger systems with AI inside. A student works through a real Matura (Polish secondary school leaving exam) paper, a language model grades the open-ended answers against the official scoring rules, and the platform turns that into a map of what the student doesn't know yet.
- Before
scope
scope
scope
Exam-prep apps handle closed questions fine, because there it's enough to compare a letter. The trouble starts with open-ended tasks, and those are what decide a biology score. To check them, someone has to read the student's answer and compare it with the scoring rules.
The result: the student completes a paper, gets a percentage and has no idea why they lost a point. The teacher knows, but can't check thirty papers a week.
A language model asked outright to grade an answer is too polite. It awards a point for an answer that sounds reasonable even when it lacks an element the key requires. For learning, that's worse than no grade at all, because it reinforces the mistake.
The second difficulty is repeatability. The same answer must get the same score today and a week from now, otherwise progress can't be measured.
The system is a modular monolith in Symfony with API Platform and a React front end. Instead of one big application we have six domain modules with separate boundaries: content and papers, solving sessions, grading, student progress, accounts, payments.
Grading is a separate module precisely because it's the riskiest part. It can be swapped out, or a second model added for comparison, without touching the rest.
The model isn't asked "how many points". It gets the task, the student's answer and the itemized key, and its job is to indicate which key elements appear in the answer and which are missing. The points are counted by code, not by the model.
That's how the student gets a sentence like "you have the mechanism, you're missing the effect on the organism" instead of a bare score. That is the whole value of this platform.
Before grading goes out to students, we test it on answers with a known score and compare it with a human grader. As long as they diverge, we don't show it as a grade, only as a hint.
We take the same approach with clients. AI gets a narrow task with a clear key, and the decision is computed by code that can be verified.
goal
goal
goal
- What's next
First, grading that agrees with a human closely enough to be shown as a grade, then a first test group of students. Only after that: accounts, payments and a public release.
- Stack
- Have a similar process?
We'll get back within 4 business hours with questions, or straight away with a proposal for how to wire it up. No strings attached.
CASE 01 · INTEGRATIONS + APPLICATIONS
Polish wholesaler · 4 price lists · 1,293 products
From paper and Excel to a single source of data: a sales rep app for tablets, a B2B portal, an LLM chatbot and an integration hub for Subiekt nexo.
CASE 02 · APPLICATIONS
Allegro listings with margin calculated before publishing
A React + Symfony application straight on the Allegro REST API, no middleman. A custom listing workflow built on a variant catalog and a profitability module.
CASE 07 · CLIENT ZERO
GitLab · Discord · n8n · CRM · helpdesk · AI reviews
We test on ourselves first. Code, clients, tickets and backups wired together with bots so a small team works like a bigger one.
- Let's start
A 30-minute call, zero cost. Tell us where you lose hours and we'll come back with an estimate of how many you can get back.
PRODAUT Sp. z o.o. · Sułków 36, 48-120 Baborów · District Court in Opole, 8th Commercial Division of the National Court Register · KRS 0001242863 · NIP 7481597304 · REGON 544821885 · Share capital PLN 5,000