Junior ML / LLM Engineer · HSE MSc in Machine Learning (with honours)
I build ML and LLM applications end to end: data, training, evaluation and a service around the model. My background combines machine learning and finance: MSc in Machine Learning at HSE and MSc in Financial Economics at HSE ICEF. Most of it is public: a self-updating radar of the Russian IT job market with a fine-tuned Qwen3-1.7B, and a retrieval benchmark with a fine-tuned embedding model for Russian law, models and datasets on Hugging Face. Open to remote roles.
| Project | What it is |
|---|---|
| Anime Video Super-Resolution | Master's thesis: ×2 anime video upscaling; RRDBNet GAN, Swin2SR and ESRT compared under three degradation pipelines, TensorRT video inference |
| ru-it-jobs-radar | Daily radar of the Russian IT job market: collectors for 58 Telegram channels and three job portals, Redpanda, ClickHouse, dbt (SCD2) and Airflow. A Qwen3-1.7B fine-tuned with QLoRA on labels from two 8B teachers extracts eight vacancy fields: macro F1 0.842 vs 0.647 for rules on held-out Telegram posts. Salary model with conformal intervals, double ML skill effects, a LangGraph text-to-SQL agent, FastAPI, MCP, MLflow, Grafana, Superset; the site rebuilds itself daily on GitHub Actions. Model, dataset and demo on Hugging Face |
| ru-law-retrieval | Open retrieval benchmark over 3,786 articles of six Russian codes in MTEB format, plus a fine-tuned embedder. 19 minutes of training on one RTX 3070 moves multilingual-e5-small from 0.802 to 0.859 nDCG@10, matching multilingual-e5-large at 4.8x the parameters and 6.8x the CPU latency. Model and dataset on Hugging Face |
| tk-rf-rag | RAG Q&A over the Russian Labor Code, now running the fine-tuned embedder above, Qdrant / FAISS behind one interface, FastAPI, answers cite articles and refuse when the context has no answer; 30-question eval with Hit@k / MRR, LLM-as-a-judge faithfulness and a chunk-size experiment. Swapping in the fine-tuned embedder moved MRR@5 on a 716-question hard set from 0.688 to 0.793 |
| vk_claude_bot | Claude Agent SDK assistant in VK chats with a tool whitelist, SSRF protection and a resource-capped Docker deployment |
| co2_detector | Battery-powered CO₂ / temperature / humidity monitor built end to end: ESP32-S3 + SCD41 + e-Paper, power-optimised C++ firmware (PlatformIO) with native unit tests, 3D-printed OpenSCAD case |
| todaysmenu | Telegram recipe bot: Qwen2.5 query rewriting, allrecipes.com scraping, streaming, LLM-as-a-judge benchmark |
| ASR with PyTorch | DeepSpeech2-style speech recognition on LibriSpeech with KenLM beam search; WER 24.4 on test-other |
| ai_text_detector | FastAPI service and Telegram bot for paragraph-level AI-text detection with a pretrained DeBERTa classifier |
| respondo | Prototype browser extension that suggests chat replies via a FastAPI backend and Gemini |
Merged:
- MLflow: #25260 Flush the
SpanBatcherfrom the trace flush paths so UC-table spans are exported - MLflow: #25514 Reject unsupported
env_managervalues when building a model container
More PRs under review in Prefect, ZenML and Feast.
Python · SQL · PyTorch · Hugging Face Transformers · LLM fine-tuning (QLoRA, unsloth) · vLLM · llama.cpp · LangGraph · MCP · sentence-transformers · embedding fine-tuning (contrastive learning, knowledge distillation) · RAG (Qdrant, FAISS) · MTEB · scikit-learn · ClickHouse · dbt · Airflow · Kafka (Redpanda) · MLflow · Evidently · Prometheus · Grafana · Superset · FastAPI · Docker · GitHub Actions · Hydra · Weights & Biases · TensorRT · LLM APIs (Claude Agent SDK, Gemini)
- Telegram: @pavel0420
- Email: alekceevpavel@mail.ru