Data Scientist / ML Engineer building large-scale recommendation & ranking systems, LLM-powered applications, and real-time multimodal pipelines.
From research and algorithm design to distributed training on 8×H100 and production integrations in C++/microservices. Also known as a Code Vibes Fixer ✨
- Recommendation & Ranking. Broadmatch recommender (LLM agents + DSSM + PMI/ALS/LightFM/TF-IDF + CatBoost), personalized recommendations for tens of millions of users, final ranking with large-scale linear models and CatBoost (C++ production).
- Scale & Performance. PySpark pipelines over 10B+ rows, distributed Transformer training on 8×H100, ONNX inference optimization.
- NLP · CV · Speech. RAG pipelines with LLM judges; real-time CV (FullHD@60fps, multi-camera trajectory/speed, face ID); real-time ASR call-bots for Meet/Zoom (no GPU) using ONNX + MediaPipe.
- Infra & MLOps. GCP, microservices, Airflow, MLflow, observability (logging/monitoring), Bazel (C++ + Python), CI/CD, tests & benchmarks.
Python |
PyTorch |
TensorFlow |
ONNX |
OpenCV |
Spark |
Airflow |
GCP |
Docker |
Bazel |
Linux |
FastAPI |
C++ |
Java |
MLflow |
MediaPipe |
- Broadmatch in search: hybrid of LLM agents, production microservices, and PMI for complementary/substitute items.
- LLM-based offline metric for similarity/complementarity evaluation.
- DSSM Transformer for semantic retrieval; PySpark pipelines over 10B+ rows.
- Distributed training on 8×H100; integrated image embeddings.
- Prod 8000 uction ONNX deployment; team CI/CD (linters, typing, tests).
- RAG microservices on Google Cloud with centralized logging/monitoring and LLM judge validation.
- Email notifications, linters, CI/CD for auto-deploy.
- Call-bot for Google Meet/Zoom: real-time A/V processing without GPU (ONNX + MediaPipe).
- Bazel monorepo (C++ + Python), unit tests and benchmarks.
- Personalized recommendations for tens of millions of users.
- Baselines: PMI and probabilistic models; production algorithms: ALS, LightFM + TF-IDF.
- End-to-end ML lifecycle in Airflow, validation metrics, MLflow tracking.
- Final ranking with a large-scale linear model; production integration into C++ services.
- Airflow pipeline from raw data to binary model; evaluation metrics & monitoring.
- CatBoost models deployed to production.
- Real-time video pipeline FullHD@60fps from two cameras.
- Ball detection & trajectory tracking, serve speed estimation, impact zone detection; face recognition on match photos.
Many repositories are private; below are public write-ups & placeholders.
- 🧠 Broadmatch Recommender — Research Notes — LLM agents, DSSM, PMI/ALS/LightFM/TF-IDF (write-up repo placeholder)
- 📚 RAG with LLM Judge — GCP, microservices, automatic validation (write-up repo placeholder)
- 📞 CallBot: Meet/Zoom Real-Time — ONNX ASR, MediaPipe streaming (write-up repo placeholder)
- 🎾 Sports CV Pipeline — 60fps, trajectory & speed, impact zones, face ID (write-up repo placeholder)
Want details? Ping me on Telegram — happy to share sanitized diagrams/code.
- 📧 Email: vkrinat574@yandex.ru
- 💬 Telegram: @Demontego
- 🌐 VK: dem0nteg0
- 💼 LinkedIn: linkedin.com/in/demontego
“Data Scientist, Researcher, Developer — and a proud Code Vibes Fixer.”