CV
Machine Learning Engineer · 6+ years of experience · Seattle, WA
Recommendation systems, MLOps infrastructure, applied deep learning, and generative AI.
summary
ML engineer with experience owning projects end to end, from cloud infrastructure and model training to production deployment. I have delivered production work across MLOps infrastructure, recommendation systems, and applied generative AI research, primarily in small, cross-functional teams.
Core stack: PyTorch, Hugging Face Transformers, FSDP, PEFT, ONNX Runtime, C++, LightFM, CatBoost, Apache Spark, Scala, AWS, FastAPI, Redis, Celery, Airflow, and MLflow.
- Set up production AWS infrastructure for Magical Labs, Media Instinct Group, and BandM8, covering API hosting and ML-pipeline scheduling.
- Took a hybrid recommendation model through to production, closing the cold-start gap for a marketplace with 200K users.
- Turned an MVP dashboard for one client into a 2+ year, multi-project relationship.
experience
Magical Labs
- Partner with founders and senior client executives to scope projects, own ML delivery, and work with backend engineers, product managers, and specialists on teams of 2–3 people.
- Design and deploy hybrid recommendation systems combining retrieval, ranking, collaborative filtering, and content-based personalization, with a focus on cold-start users and items.
- Build and maintain production ML infrastructure on AWS, including EC2, Elasticsearch, Redis, PostgreSQL, FastAPI, Airflow, and MLflow.
- Reproduce and benchmark three research papers on high-capacity image steganography with latent diffusion models; optimize parameters with Weights & Biases Sweeps and improve robustness to heavy JPEG compression.
- Deliver client-facing data products, mentor junior engineers, and contribute to technical hiring.
BandM8
- Took an AI drum-generation feature from early experiments to production code shipped to users.
- Fine-tuned GPT-2 for drum-pattern generation (Tap2Drums) with multi-GPU FSDP training and parameter-efficient fine-tuning; refined the MIDI tokenizer for drum-specific patterns.
- Proposed audio-transient features to convert voice recordings into structured MIDI input, enabling drum-pattern generation from a hum or vocal recording.
- Built the pipeline from raw MIDI collection to training-ready datasets; exported the model to ONNX and rewrote inference in C++ for use in Logic Pro and Ableton.
Media Instinct Group
- Rewrote and tuned Python Spark jobs in Scala, transforming advertising and session logs into clean training data reused across the team’s modeling efforts.
- Set up MLflow on AWS for experiment tracking and Apache Airflow for job scheduling, giving the team a shared process for comparing model versions.
- Built and tuned a CatBoost ad-ranking model end to end, using user-session context, creative metadata, and historical CTR signals.
MaximaTelecom JSC
- Owned exploratory analysis, feature engineering, model training, and evaluation for advertising and personal-finance products.
- Developed an 80-category NLP personal-expense classifier and a rule-based system for extracting product and brand names; both shipped in the expense-management product.
- Built an embedding-based item recommendation system with ANN candidate retrieval, using classified expense history and item-interaction signals.
- Explored early-stage ideas through feature-engineering experiments and proof-of-concept builds to help prioritize deeper investment.
education
Lomonosov Moscow State University
Lomonosov Moscow State University
skills
ML & research
PyTorch, Hugging Face Transformers, FSDP, PEFT, ONNX Runtime, LightFM, CatBoost, BERT
Data & infrastructure
Apache Spark, Scala, AWS, FastAPI, Redis, Celery, Airflow, MLflow, Elasticsearch, PostgreSQL
Languages
Python, C++, Scala, SQL, Bash
Interests
Mechanistic interpretability, cognitive modeling, generative AI research, and steganography
certifications
- Create Docker Container with Flask
- Seaborn Regression Plot App
- Deploy Machine Learning Model into AWS Cloud Servers
- Google Data Analytics Certificate