Tsinghua University — AI Researcher (Aug 2024 - Oct 2025)

Multimodal fusion robustness: Enhanced multimodal information processing systems by integrating weighted fusion networks and modality reconstruction pipelines (~2.7% accuracy improvement).

Repository documentation generation: Built and deployed a multi-agent MVP (AutoGen, FastAPI, Docker) that auto-generates documentation for code repositories.

VTB Bank — ML Engineer (Sep 2023 - Aug 2024)

Proactive anomaly detection: Built a time-series anomaly detection pipeline (isolation forest, Prophet, volatility analysis) deployed in production (Airflow, Docker, PostgreSQL), enabling identification of system failures hours ahead of human monitoring. In addition, implemented a time series classifier across 50+ features (sliding window statistics, simple time series features, changepoints, correlations, etc), and integrated it into an anomaly detection pipeline.

Automated support request classification: Built a text processing pipeline for the technical support service to automatically classify tickets based on request title and body: TF-IDF + SVD + mini-batch K-Means for unsupervised request clustering, followed by a LightGBM classifier for automatic ticket routing - reducing average processing time by ~25%.

Sberbank — Data Scientist (Mar 2022 - Sep 2023)

Demand forecasting: Implemented a production demand forecasting system across 100+ time series (product × supply center × customer triplets). Generated lag, product, and time features with Prophet and linear regression predictions as meta-features fed into a CatBoost model (Hyperopt-tuned), improving supply efficiency by ~20%.

Geo-spatial store location optimization: Built geo-spatial ML pipelines to identify optimal store locations and forecast company revenue for multi-format retail expansion across 15+ business cases, using spatial feature engineering and big data processing tools.