Knowledge distillation of SOTA encoder transformers into state-space and efficient-attention architectures at VK, supervised by the author of DenseAttention (ICML 2025). Also: CTO of an AI-coaching startup with a production iOS app, and founder of LooksAR — an AI virtual try-on widget for optical e-commerce (Three.js + MediaPipe). Previously: model validation at Sberbank (risk), DS for a large marketplace.
| Competition | Result | Approach |
|---|---|---|
| NeurIPS 2025 Ariel Data Challenge | 🥈 31st · Silver | Exoplanet spectra from noisy telescope time series — Gaussian processes, Bayesian inference, physics-informed modeling, no neural nets |
| Santa 2025 (Kaggle) | 🥈 108th of 3,400 — top 3% | Combinatorial optimization — custom C++/Rust packing solvers |
| Data Fusion 2026 «Guardian» | 🎖 11th / 574 | Anti-fraud on 200M+ banking transactions, extreme class imbalance |
| Alfa Bank: Predict Next Transactions | 🎖 6th | Multi-label transaction sequence classification |
| Data Feeling RAG Challenge | 🥇 1st | Hybrid retrieval: BM25 + dense encoders |
| Wunder Fund LOB Predictorium | 51st | Limit order book prediction on HFT data — GRU + MLP ensemble |
| AI Academy Molecule Hackathon (ODS) | 🥉 3rd / 51 teams | Optical chemical structure recognition (image → SMILES), ₽1M prize pool |
timeline
2023 : ML freelance — CV/OCR pipelines, marketplace pricing
2024 : Sberbank — model validation (risk) : LZT Market — marketplace DS
2025 : NeurIPS 2025 Ariel — 31st, Silver : CTO of an AI-coaching startup — present : 2 iOS apps shipped solo
2026 : AI Researcher @ VK — present : Founded LooksAR — AI eyeglass try-on : Santa 2025 — top 3% : Data Fusion «Guardian» — 11th of 574
| Repo | What it is |
|---|---|
| sparrow-santa2025 | Rust 2D polygon packing optimizer — Santa 2025, top 3% |
| Molecule-hack-3d-place-solution | 3rd place OCSR pipeline: MolScribe cascade + RDKit validation |
| steam_account_price_prediction | Production pricing models with ONNX deployment (R² 0.92) |
| trade_comrade | Automated crypto trading research system on Bybit |
| kaggle-kernel-tools | Claude Code skills for remote Kaggle kernels |
| dense-attention | Contributions to DenseAttention / DANet (ICML 2025) |
📍 Moscow · ✉️ adakmaev@edu.hse.ru · 💬 t.me/DoSwyazi · 🏅 kaggle.com/oostg0t

