YC Startup School 2026

2026-05-09• Lausanne

How I Got Into YC Startup School 2026

Somewhere I saw that YC opened applications for YC Startup School 2026. I applied on Saturday, March 28 at 12:59 PM and received an acceptance letter from Diana Hu on Friday, April 17 at 12:59 PM. I'm extremely excited about this opportunity, it will be life-changing.

“The journey of a thousand miles begins with a single step.” Golden Ticket

Here, I just want to save my responses for my future self, so I can look back and reflect on this journey:

List any competitions/awards you have won, or papers you’ve published.

  1. https://x.com/arcprize/status/1971587588386611320

ARC-AGI 2 (Kaggle 2025) Top 2 / 1456 - Co-developed a reasoning model that placed 2nd out of 1,456 teams (27.08% score), outperforming models such as GPT-5 Pro (18.3%) and Grok-4 (16%).

  1. https://www.imo-official.org/participant_r.aspx?id=26026

International Mathematical Olympiad (IMO 2018) — Bronze Medal

  1. https://www.imc-math.org.uk/?act=results&by=sum&year=2022&onl=3#:~:text=Mekhron,Bobokhonov

International Mathematics Competition for University Students (IMC 2022) — Gold Medal.

Tell us about things you've built before. For example apps you've built, websites, open source contributions. Include URLs if possible.

    1. ARC-AGI Reasoning Systems (Kaggle 2024–2025): https://x.com/arcprize/status/1971587588386611320

Co-developed a reasoning model that placed 2nd out of 1,456 teams on ARC-AGI 2 (27.08% score). Built large-scale training pipelines (up to 128×H100 GPUs), implemented custom architectures and optimizations (FlashAttention, quantization, CUDA kernels), and worked across pretraining and RL-based post-training. At the time of the competition, this outperformed reported results from frontier models such as GPT-5 Pro (18.3%) and Grok-4 (16%). Also designed a test-time training framework enabling models to improve during inference, increasing ARC-AGI-1 performance from 11% → 35% (+24 pts).

  1. Large-scale NLP Safety Systems (Yandex Search)

Built and deployed production ML systems for harmful content detection, serving ~50M daily users. Reduced harmful results for suicide-related queries by 92%, and developed autocomplete filtering models (88% recall / 86% precision) at scale.

  1. FindOFF — https://mekhronbobokhonov.com/projects/findoff/

Built a web app that aggregates and ranks ML/software engineering jobs in Switzerland, with filtering, tracking, and daily updates.

  1. Personal Website — https://mekhronbobokhonov.com

Built and maintain a personal site to showcase projects and research.

Please tell us about a time you most successfully hacked some (non-computer) system to your advantage.

When I was looking for an apartment in Switzerland, I found a place I liked but was told I was the 5th person to view it and that the real estate company would decide who gets it.

I realized the queue wasn’t the real decision mechanism, the outcome depended on the property manager’s perception of risk. Instead of waiting, I went directly to the property management office and asked to speak with the person responsible.

I presented my work contract, proof of financial stability, and marriage certificate to show I was a reliable tenant. I also created urgency by telling them I needed a decision immediately, otherwise I would move on.

By going straight to the decision-maker and reducing their uncertainty, I was able to bypass the queue and secure the apartment on the spot.