I focus on the intersection of applied ML and ML Engineering — from experiments to deployment.

Currently: polishing projects, diving deep into MLOps, and open to Data Scientist / ML Engineer positions.

ML / DL:

Data:

Engineering:

1. Music Genre Classifier:

   Group project — ML lead contributor

   • Designed and trained a 2 MB CNN, achieving F1 macro = 0.82 without data leakage; accuracy is comparable to a 380 MB fine-tuned transformers
   • Adapted the model for inference in a container using Docker and RabbitMQ (The container was configured by the other member)
   • Team practices: PR reviews, conventional commits, feature-branch flow, CI checks on all merges

2. Support ticket prioritization:

   End-to-end ML project — classification on imbalanced data with business constraints

   • Built a priority classifier (Low / Medium / High) on a 50K synthetic dataset with severe class imbalance
   • Eliminated critical errors: zero cases of High→Low misclassification
   • Used TreeSHAP to verify model relies on real business factors (downtime, affected users)

3. Ultrasonic welding optimization:

   R&D thesis project for the "Polymer Composite Materials" lab, SPbPU
   Bachelor's thesis — defended with excellence 25 June 2025 🔒 private — refactoring in progress 🔒

   • Implemented a surrogate assisted bayesian optimization system based on CatBoost + Optuna
   • Achieved 1.4–3× reduction in the number of experiments needed to find the global optimum compared to classical Bayesian optimization
   • Validated on benchmark functions (Branin, Rosenbrock, harmonic) with systematic methodology
   • Compared:
     • CV configurations (4-,6-,8-fold and LOO)
     • ML models for a surrogate (CatBoost, XGBoost, RandomForest, MLP with Nadaraya-Watson kernel)
     • Optimization methods (CMA-ES, TPE, GP, QMC)

4. Topological Data Analysis research:

   Research group project: assisted with TDA pipeline; responsible for the classifier

   • Built a TDA pipeline extracting robust geometric features from 3D point clouds using gudhi and giotto-tda
   • Improved F1-macro from 0.73 to 0.81 (+10%) over the baseline CNN, demonstrating effectiveness of topological features