Asgat Akmaev

Distilled a large encoder model into a compact DenseAttention model on LRA ListOps. The student with 1.4M parameters outperformed a vanilla Transformer with 19.2M parameters by 4% in accuracy. Implemented knowledge distillation for a BERT classifier and optimized training on a GPU cluster. Investigating state-space architectures for encoder-model distillation, supervised by the author of DenseAtte

Technical co-founder of dotproject.app, an AI-coaching startup. Own the entire technical side: production iOS app (SwiftUI), backend, payments infrastructure (T-Bank acquiring), analytics. Shipped to the App Store.

• Built and deployed an anti-fraud system from scratch as part of a small team. The solution was accepted and paid for by the client.
• Investigated data, engineered features and tested hypotheses. Trained a quantile regression model for price estimation and developed a fraud detection heuristic. Precision@300 on compliance-labeled cases: 5.28%, compared with 0.43% for a price-based rule.
• Built

• Independently validated risk models before production deployment through backtesting, assumption analysis and robustness checks across market regimes. Prepared validation documentation.
• Approved an LTV model for production and returned two other models for revision with documented findings.
• Built an IT-validation tool for an internal model-validation agent.

Grade: 7.5

Activities and societies: Study

Coursework: probability theory and mathematical statistics, mathematical analysis, linear algebra, ordinary differential equations, complex analysis, functional analysis, numerical methods and C++.

Pre-university achievement: 4th place nationwide in the National Technology Olympiad, Artificial Intelligence track (team competition).

GPA 7.45