Regional Championship in Competitive Programming

The company I work for occasionally offers interesting activities - hackathons, conferences, thesis defenses, olympiads. Not as a participant, but in the role of an expert.

This time the activity was the Lipetsk Region Championship in Product Programming. My task was to judge and prepare the challenge. I had an example from last year, so I roughly understood the difficulty level.

According to the regulations, teams had about 2.5 days to complete the task, but considering modern technologies (primarily AI), the challenge was both challenging and feasible at the same time.

I won't fully disclose the challenge, but it boiled down to the following:

• a personal fitness assistant;
• preparation of an exercise plan and nutrition program based on user parameters and preferences;
• ability to make adjustments through chat;
• and yes, of course, all based on AI, with RAG and all that.

Until the last moment, I didn't know either the teams or their projects - access to the code (as I was the technical expert responsible for code evaluation) was provided on the morning of the third day, when there were only 4 hours left before the presentations began.

There turned out to be 9 teams, each with 3-5 people, meaning I needed to review the code of approximately 36 people in this time. Many projects were microservice-based, which didn't simplify the task.

But if participants can use all tools available, why should judges be different? My evaluation pipeline was as follows:

1. Clone all projects and subprojects into a folder on my laptop.
2. Open everything as a project in Cursor.
3. Write a bash script to go through all folders and pull updates (as work was still ongoing and participants were pushing changes on the fly) - and git wasn't straightforward either, besides multi-repositories there were also submodules. One team stood out - they had a single repository where backend and frontend were in different branches - where else would you see that?
4. Write a prompt for an assistant to generate a report on each project - listing important technical details and a clear report structure - a separate markdown file is generated for each team.
5. With a separate prompt, generate a summary report from the detailed reports, with tables and scores for specific criteria for each team.
6. If the update script pulls changes - regenerate the team's report and the summary.

Thus, by the start of presentations, I had all the data on the projects, ranging from database schemas and used models to system prompts and even exposed API keys (yes, one freshman team committed their API key).

Results

It was a surprise that all teams had to complete the same task (yes, the one I designed) - this is the regulation for competitive programming. In hackathons, there are usually several tasks to choose from.

Out of 10 registered teams, 9 made it to the finals. Only 2 were notably weak - both freshman teams, and I believe they have their future ahead of them.

The top three teams created nearly finished products that could be polished and sold - there were no analogues on the market, at least in Russia.

Only one team implemented honest RAG with embeddings - and they won.

In the challenge, I also mentioned structured output, but everyone implemented it through prompting - simply asking for JSON in the response. By the way, this is a feature of some vendors and models (OpenAI at least) - when used, the response structure is 100% guaranteed. The teams probably didn't know about this, and the feature isn't widely available across models, especially those accessible within the country.

Overall, there's a complete feeling of accelerated technological progress - 10 years ago, across my entire university cohort (studying the most programming-oriented specialization), I didn't know anyone who could create something like this, especially in such a timeframe. And it's not just about the AI needed to implement the tasks, but about AI assistants that exponentially accelerate research and implementation of new projects, especially concepts.

It's still difficult to implement something complex and large-scale with AI without personal immersion in the field, but that's only for now. I'm confident we're just at the beginning of the journey, and no one can predict what will happen in 10 years. As a result, I'll now have a judging category.

The university website published a news article about the award ceremony, and I even appeared in the group photo.