B.Sc., Applied Mathematics and Computer Science
Saint Petersburg State University
2027Mikhail Sharabarin
@mike-sharabarinConnected with DSML. This person's Telegram account has been confirmed through DSML Kazakhstan.
Machine Learning Engineer | Data Scientist
I like digging into what changes when you quantize a language model, and I also enjoy taking ML out of the notebook. One recent project got ECG predictions running on a Raspberry Pi 5.
Open to
Hiring & referralsSpeaking & workshopsResearch collaborationPaper discussions
Career
Education
Selected projects
LLM KV-Cache Quantization
Studied FP8 and HQQ INT4/INT2 quantization on three reasoning models and 80 AIME-24 and MATH-500 tasks. Developed analyses to investigate reasoning degradation and found architecture-dependent quality loss and quantization error in outlier channels.
Open projectAutoTopic, Topic Modeling
Built a multilingual BERTopic pipeline and compared it with LDA, using Optuna for hyperparameter optimization. Identified 59 topics from 236K documents and achieved diversity of 0.966, compared with approximately 0.91 for LDA.
Open projectDistributed ML Inference
Implemented 50-topic classification and a distributed inference system using a FastAPI coordinator, gRPC workers, Kubernetes, Cassandra, and MinIO. Scaling from one to five worker replicas raised throughput from 87.1 to 346.2 requests per second and reduced p50 latency from 684 to 165 ms.
Open projectECG Edge AI
Built an ECG acquisition and inference system using two AD8232 modules, two Arduino Nano boards, and a Raspberry Pi 5. Implemented six-lead reconstruction, filtering, a signal-quality gate, and CPU-only inference at 5.18 ms per prediction.
Open projectAchievements
Top 1%, Gradient of Growth time-series forecasting competition
Top 15%, MTS Challenger
Skills
Classic MLDeep LearningNLPLLMsRecSysDiffusion ModelsPyTorchPythonC++SQLDockerGitReactNext.js