The client’s product is an international educational platform for self-learning in mathematics and physics, for completing courses and preparing for Unified Exams. Previously, methodologists manually created training and assessment tasks in their admin panel, which was time-consuming and sometimes lacked a creative approach.

Our task was to implement AI that would simplify the methodologists’ work by automatically generating various types of tasks based on a request, as well as being able to shorten, simplify, and translate them into other languages.

Possible task generation scenarios:

• Duplicate an existing task, but with modifications (for example, with different numbers);
• Create new types of tasks based on the old ones.

We conducted research and compared popular AI models to identify the most effective one, to assess differences in generation using the same approach, as well as to gather information on possible limitations.

The Gemini 2.0 model yielded the best results; we even asked the AI to write its own prompt based on the specified requirements and conditions. All the generated tasks were correct, but as part of our research we also developed a validator that checks the tasks for solvability, the conformity of the answer to the task, and the consistency of intermediate answers with the corresponding intermediate steps.

In the end, we produced a universal and endless turnkey task generator.

It creates task scenarios and outlines the entire solution algorithm with step-by-step explanations. Now the methodologist only needs to log into the admin panel, navigate to the LLM interface, paste the prepared prompt, then download and import the generated result into the admin system.

• A working MVP integrated into the client’s infrastructure.
• The methodologists’ workflow has been automated and the company’s expenses have been reduced.
• The tasks for students have become more engaging and creative.