Mekhron Bobokhonov
AI researcher / engineer with experience on large-scale reasoning models, distributed pre-training, RL-based post-training and test-time adaptation. Contributor to sota ARC-AGI systems.
Interests: Training/Optimizing/Debugging LLMs/RL, Startups, Entrepreneurship, Personal Growth/Finance.
Competitions: ARC-AGI 2 (Kaggle 2025) Top 2 - Gold Medal , ARC-AGI 1 (Kaggle 2024) Top 13/0.8% Gold Medal, IMC 2022 Gold Medal, IMC 2021 Silver Medal, IMO 2018 Bronze Medal.
AI Research Scientist
Giotto AI
Lausanne, Switzerland
• Contributed to training large-scale reasoning models across pre-training and RL-based post-training.
• Designed and ran distributed training pipelines (up to 128×H100), implementing custom model architectures and low-level optimizations (FlashAttention, quantization, custom CUDA kernels).
• Co-developed a reasoning model achieving Top 2 / 1456 teams on ARC-AGI 2 (Kaggle 2025) with 27.08% (GPT-5 Pro achieved 18.3% and Grok-4 16%). Technical Report
• Designed and implemented a test-time training framework enabling online adaptation during inference; Improved ARC-AGI-1 (Kaggle 2024) benchmark performance from 11% to 35% (+24%), contributing to Giotto.ai's Gold Medal, Top 0.8% globally (1,428 teams).
+Middle Machine Learning Engineer
Yandex
Remote
• Worked in the Special Search Projects team — an internal machine learning consultancy within Yandex.Search that helped other teams develop and deploy ML-driven solutions they couldn't build in-house.
• Built and deployed NLP models for harmful/sensitive content detection in Yandex.Search, improving safety for high-risk query categories.
• Achieved a 92% reduction in harmful material for suicide-related queries and significantly improving user safety.
• Developed an autocomplete filtering model (88% recall / 86% precision) for detecting inappropriate or self-harm-related
suggestions at scale.
• Contributed to the design and implementation of a large-scale data collection pipeline for training a Perplexity-style context-aware search model.
Master's degree in Data Science
EPFL
Lausanne, Switzerland
• I combined full-time job at Yandex with master's studies. GPA: 5.25 / 6.0.
• Took advanced courses in Machine Learning, Natural Language Processing, Computer Vision, Reinforcement Learning, and related topics.
• Conducted research projects in the field of Computer Vision to gain hands-on experience, focusing on unsupervised and open-vocabulary semantic segmentation.
• Completed my master's thesis at AXA Group Operations, where I worked on transferring state-of-the-art open-vocabulary segmentation models from the natural image domain to the aerial domain.
Data Science Intern
Yandex
Moscow, Russia
• Worked in the anti-fraud team of Yandex.Search.
• Developed and deployed bot-detection models that reduced automated traffic load on Yandex Search, improving system efficiency and resource utilization.
• The model successfully detected and prevented over 4 million robots in Yandex search.
Data Science Master's Program
Yandex School of Data Analysis
Moscow, Russia
• The most prestigious and selective Master's-level Machine Learning private study program in the post-Soviet region (3% acceptance rate — 129 students admitted out of 4,300 applicants).
• Coursework included: Machine Learning, Natural Language Processing (NLP), Computer Vision (CV), Reinforcement Learning (RL), and related topics.
• Studied in parallel with my Bachelor's degree.
Bachelor of Applied Mathematics and Computer Science
Moscow Institute of Physics and Technology
Moscow, Russia
• The MIT of Russia. GPA: 4.77 / 5.0, Diploma with honours.
• Coursework included: Algorithms and Data Structures, Probability Theory and Statistics, Calculus, Linear Algebra, Topology, and Discrete Mathematics.