X5 Tech explains how a team built an AI service for international exam prep in 7 days

During the AI Talent Camp intensive, the ExamLab Bot team spent a week building a working AI service for tutors who teach international exams. The system turns a syllabus and deadlines into a personalized learning path and cuts plan preparation from 3–4 hours to roughly two minutes.

What was the problem?

The idea grew out of a very specific pain point for IB and A-Level tutors. To take on a new student, a tutor can't just read the syllabus. They have to turn it into a detailed 20–30-week plan that accounts for the exam deadline, the schedule, topics already covered, and the student's current level. If a lesson gets rescheduled or the student falls behind, the plan has to be recalculated all over again. Each student takes about four hours of routine prep, and it is this work, not the teaching itself, that caps growth.

As a result, many tutors don't maintain proper personalized learning paths at all. Some go by intuition, some use one template for everyone, and some simply turn away new students. In a segment where lessons with a strong expert cost 30–40 thousand rubles a month, that translates directly into lost revenue. The ExamLab Bot team set out to automate exactly this repetitive organizational layer: building the plan, accounting for constraints, updating it based on progress, and tracking movement toward the goal.

How the MVP works

With only seven days, the team deliberately narrowed the scope to two core functions: building a personalized learning path for a student's goal, timeline, and level, then adjusting it as the student progresses and the tutor gives feedback. They chose a Telegram bot as the interface to avoid spending time on a heavy web frontend. Google Sheets became the working view of the plan, and the schedule syncs through Google Calendar. Instead of 3–4 hours of manual prep, generating a plan now takes about two minutes.

• the tutor or administrator sets the goal, deadline, level, available time, and priorities
• the bot generates a detailed study plan for the entire period
• the plan is saved to Google Sheets, and events go to Google Calendar
• when progress changes, the AI recalculates the learning path without rebuilding it from scratch

Under the hood, the service runs on a multi-agent architecture. A Python orchestrator coordinates agents with different roles: one handles strategy and time allocation, a second fleshes out lessons, homework, and tests, a third batch-processes long syllabi, and a separate layer validates structure, dates, and coverage completeness. The stack includes Python 3.11+, asyncio, aiogram, SQLAlchemy, PostgreSQL, and the OpenRouter API. The team skipped LangChain and LangGraph in favor of direct calls, which made it easier to control prompts, retry logic, and development speed.

"A working minimal product with a clean pipeline is worth more than

trying to do everything at once."

Why did they make it in a week?

The intensive itself was structured as a short product cycle, not a typical hackathon where the demo is the point. Day by day, the team moved from discovery and problem definition through risk assessment and a PoC to an MVP, user feedback, and a final presentation. Mentors kept reminding participants that the goal wasn't a polished stage show but groundwork for a real AI product that could keep evolving. That pushed the team to make architecture decisions quickly, drop unnecessary features, and stick to a working scenario rather than flashy ideas.

The project's authors specifically point out that the speed didn't come from the magic of AI tools alone. The interface, the tutor interaction flows, and even the choice of the first user segment changed several times over the week. Betting on a minimal working version and fast checks against real scenarios proved more useful than trying to design a perfect system up front. The next step is a closed beta with 10–15 tutors, followed by more exams and support for Unified State Exam (EGE) prep.

What does it mean?

The ExamLab Bot case shows that in EdTech right now, narrow AI products with a clear value metric work best. Here, that metric was time: hours of manual prep became minutes, and tutors got a chance to scale their practice without taking on more administrative work. There's an important lesson for other teams, too: simple interfaces, direct integrations, and frequent user testing often deliver more than a complex stack and a long feature list.