Fourth-year student at Bauman Moscow State Technical University (Dept. IU5), graduating in 2027. I build models on tabular data and time series, work on uplift modelling and experiment statistics, and also do computer vision and audio. I take models through to a service with an API, tests and Docker. Open to Data Science and ML internships.

Area	Stack
ML · Tabular
DL · CV · Audio
LLM · NLP
Data · Infra

X5 RetailHero Uplift · who actually changes behaviour when contacted  

Uplift modelling on 45.8M receipt lines in PostgreSQL (dbt, 42 tests). T-, S- and X-learners on CatBoost and the causal metrics (Qini, AUUC, uplift@k) written from scratch. The top 10% by response and the top 10% by uplift overlap by only 0.44% against 10% by chance; Δ uplift@10% = +9.53 pp [+6.63; +12.54], 2,000-replicate bootstrap. Response-based targeting loses money (0.53 pp against a 4.67 pp break-even), while uplift targeting is worth ₽612k per campaign. The analysis plan was pre-registered before the first model, and a mutation test proves the leakage auditor works.

E-CUP 2026 (Ozon) · 30-day GMV forecast for 250,000 users · team of 2  

14th of 316, private RMSLE 1.6631. Every training row is a user × date anchor with features strictly before that date, the model is trained directly in the metric's scale (log1p), and the final model is a GRU + gradient boosting ensemble. Adversarial validation exposed an anchor-date leak (AUC 1.0); once it was removed, 93% of the local improvement turned out to be illusory.

OreScope · ore grade from thin-section microscopy · Nornickel AI Science Hack finalist  

U-Net phase segmentation on panoramas of up to 300 MP; the grade follows from an explicit expert rule over the mask. macro-F1 0.91 on intergrowth type, 0.77 end-to-end. Colour normalisation lifted the unseen camera domain from 5 to 14 samples out of 20 without new labels. A panorama takes 124 s against a 5-minute requirement; FastAPI, Docker, built in 48 hours.

Stem Separator · splitting a track into four stems  

Demucs (htdemucs) behind a REST API, a web UI and a CLI. SDR 9.34 dB on vocals and 9.90 dB on bass (museval, MUSDB18-sample, 5 tracks). A 3:17 track is separated in 14 s on an M4 Pro (×14 real time). MPS/CUDA/CPU autodetect, chunked inference, 19 tests, Docker Compose.

FAD Benchmark · evaluating generative music models · research at Bauman MSTU  

Five open-source text-to-music models (MusicGen, AudioLDM-M/L, MusicLDM, Riffusion) compared with FAD, FAD-inf, per-song FAD and CLAP Score, using three embedders (CLAP-LAION-Music, MERT-v1-95M, EnCodec) and two reference sets. The leader depends on the reference: AudioLDM-M on FMA-Pop (CLAP-FAD 0.039), MusicLDM on MTG-Jamendo (0.0044).

AI Team Assistant · LLM assistant with a strict answer format  

Google Gemini behind a server-side route with structured output through responseSchema, so the JSON is valid by construction. Prompt-evals run 4 synthetic requests, including a prompt injection, and check the structure and content invariants.

Smoking Detection · detecting smoking in a video stream, on-device  

YOLO26n on person / smoke classes: mAP50 0.713, mAP50-95 0.317, precision 0.744, recall 0.680. Smoking is flagged by box overlap, and the model is exported to CoreML (.mlpackage) for inference on iOS.

Year	Competition	Result	Link
2026	E-CUP 2026 (Ozon): 30-day user GMV forecast · team O3	RMSLE 1.6631	solution
2026	Yandex ML Challenge (Young & Yandex), final		certificate
2026	Yandex School of Data Analysis: AI Agents Security Week	guardrail, from 61.54 · red team 71.43 / 100	certificate
2026	DatsSol hackathon (DatsTeam): colony-control bot		solution
2026	Nornickel AI Science Hack: ore classification from microscopy	macro-F1 0.91	solution