Math & CS student at Saint Petersburg State University · Deep Learning · RecSys · GraphML · LLM Agents · Classic ML

Building and breaking ML systems — recommenders, graph models, boosting pipelines — to find where a good metric stops being honest.

Used in projects: Python, PyTorch, PyTorch Geometric, scikit-learn, XGBoost, CatBoost, LightGBM, NumPy, Pandas, pyarrow, Optuna, SHAP, Git, Docker, FastAPI, Ollama, C++

Worked with: SQL (PostgreSQL), Kafka, Airflow, Linux, CI/CD (GitHub Actions), LangGraph

Decoder-only transformer for sequential recommendation, trained on 881K events from Yandex Music. Causal self-attention implemented from scratch; track audio embeddings are quantized into semantic IDs via RQ-VAE, turning recommendation into generative retrieval — the vocabulary collapses from 181K items to 1,026 tokens, making the model 45x lighter.

• NDCG@10 of 0.575 for the hybrid vs. 0.413 for standard item embeddings.
• Pure generative retrieval collapses (0.181) — per-position perplexity explains why: RQ-VAE was trained to reconstruct audio, so only the first of four quantization levels is predictable from behavior.
• Found a metric bug: the last-item baseline showed HitRate@10 = 1.0 due to rank ties; after a fair tie-break, it dropped to 0.0145.

Money-laundering detection on 5M transactions with a 1:1122 class imbalance. Transfer graph topology compressed into 86 features via sliding windows over prefix sums: searchsorted instead of groupby().rolling(), which doesn't fit in memory — 40 seconds for 497K accounts.

• PR-AUC of 0.348 vs. 0.011 for the naive baseline, 0.462 on holdout, Precision@100 = 0.84.
• Error analysis by scheme type: star-shaped schemes caught at recall 0.53–0.62, chain schemes at 0.25–0.32 — the boundary tracks topology, since all features are limited to a one-hop neighborhood.
• Feature-group ablation diverged from gain-based importances: the receiver side contributes twice as much as XGBoost's importances would suggest.

SO(3)-equivariant prediction of vector and tensor quantities: symmetry is built into the architecture rather than learned from augmentations. SchNet plus rank-1 and rank-2 heads, on QM9S.

• MAE of 0.038 D vs. 0.639 for a larger fully-connected network.
• Equivariance verified numerically on the untrained model.
• Three-seed ablation: topological features are useless, since they lie in the kernel of the group action.

• Saint Petersburg State University, Faculty of Mathematics and Mechanics, "Mathematics and Computer Science" — 2nd year, 2025–2029
• Deep Learning School, MIPT — 2025
• Agents Week, YSDA (Yandex) intensive — 2026

The most interesting moment in ML, for me, is when a model hits a ceiling — and it turns out the problem isn't its size, but how the object is described. Three projects in a row, I ended up at exactly that point, each time from a different angle.

Interests: graph methods, deep learning, LLM agents.

Contact: Telegram · mmworkmaildaaa@gmail.com

English, German - fluent, Russian - native