An AI agent that checks math and physics homework—implemented within 14 days for the High School Graduates School

Who is the client?

Szkoła Maturzystów offers courses to prepare students for the high school graduation exam, led by Dr. Łukasz Jarosiński, a graduate of AGH University of Science and Technology. Over the past 15 years, more than 5,000 high school seniors have completed these courses. The average score of course participants on the advanced-level math and physics exams is approximately 80%, compared to a national average of about 35%. Łukasz’s teaching system has been published in a peer-reviewed scientific journal.

What challenge needed to be addressed?

Łukasz’s method is based on homework assignments. After each lesson, students receive homework assignments designed using the interleaved method—which, according to research, is the most effective but also the most labor-intensive to implement. Homework is effective only when it is checked and discussed, and with over five thousand students, the workload associated with grading grows along with the school.

Łukasz himself referred to this in his testimonials as “the painstaking process of verifying and evaluating homework.”

What does the agent do?

We created a system for the automatic grading of high school-level math and physics homework assignments at the advanced level, using artificial intelligence models to analyze students’ solutions and generate feedback. We delivered it as an API.

The system accepts four elements as input: a photo or scan of the solution, the problem statement, the grading criteria, and the user ID.

During the analysis, the system interprets the student’s solution, evaluates its correctness, identifies errors and inconsistencies in the line of reasoning, and determines whether the solution meets the provided grading criteria. The evaluation is therefore based on criteria, not merely on the correct answer itself.

Between assignments, the system keeps track of a user’s errors: it identifies those that recur in subsequent assignments and takes the error history into account with each subsequent assessment.

As a result, it generates a response that includes an assessment of correctness, an indication of errors, information about recurring errors, and recommendations and tips for improvement.

These last two elements distinguish this solution from a simple assignment check. The student does not receive a series of independent evaluations—they receive information about the mistakes they consistently make. Typically, only the teacher, who remembers the student’s previous work, has access to this information.

What was included in the scope, and what was deliberately left out?

In addition to the system itself, we provided three sets of documentation: API documentation, integration documentation, and technical documentation covering architecture, implementation, and maintenance. We deployed the system in a production environment.

The contract clearly specifies the purpose of this documentation: it is intended to enable further development of the system by a third party. This is not a declaration of goodwill, but a written commitment—the client is not dependent on us and can develop this system with anyone at any time.

The scope did not include the user interface or client application, end-user support, or functional development after the project’s completion. Szkoła Maturzystów has its own e-learning platform and its own website developer, so there was no reason to build a second layer alongside the existing one. Our task was to develop the grading engine.

We believe that for a client who already has a mature product, this is the right approach: we step in where there is an actual lack of expertise and do not duplicate what the client has already accomplished.

How did the collaboration go?

The project took 14 calendar days to complete. The work was carried out by Mateusz Cieśliński and Jakub Dulas—directly, with no intermediaries between the client and the developers. In his testimonial, Łukasz noted that the project was completed according to the agreed-upon schedule.

The implementation is covered by a 12-month warranty, and response and bug-fixing times depend on the severity of the error—ranging from critical to minor. Maintenance includes monitoring the system’s operation, fixing errors, security updates, and a quarterly review of the AI module’s performance.

We intentionally included this last point in the contract and believe it should be standard practice for AI-based implementations. A traditional system either works or it doesn’t. An AI module can be functioning while simultaneously performing increasingly poorly—which is why its quality should be checked periodically, rather than waiting for a report from the client.

What did the client say?

“The implemented solution has significantly streamlined our work, allowing us to automate the tedious process of verifying and grading homework, saving time and improving the quality of feedback provided to students.” – Dr. Łukasz Jarosiński, High School Graduates’ School (testimonial, April 2026)

Who benefits from this model?

For companies where employees perform repetitive substantive evaluations: they read others’ work, compare it to criteria, and prepare justifications. Schools and educational platforms—but not only those—anywhere where an evaluation process exists, criteria are in place, and a reasoned response is required.

There is one condition, and it stems directly from how this system works: evaluation criteria must exist and be communicable. In this implementation, evaluation criteria are one of four input elements—without them, there is nothing to apply. If, within a company, the evaluation exists solely in one person’s mind and no one has ever written it down, the project begins with defining the criteria, not with writing code.