Programs

Streaming prediction of public transit delays: binary NDTP protocol over TCP, GPS matching with schedules, HMM mapping of tracks to OSM streets, CatBoost on 69 features with strict anti-leakage. Fair streaming evaluation: MAE 42 s vs. 93 s for the baseline. Dispatcher dashboard on MapLibre with “what-if” scenarios, Docker Compose (3 services), ONNX export, 18 tests. Hackathon solution, built solo in 2 days using agents, then pushed to the repository.

Norbert AI — a ready-to-use architecture for building intelligent assistants. Semantic search, graph, RAG, web search and scientific database search, annotation, and analytics via LLM.

Device defect inspection pipeline based on a vision-language model: fine-tuning Qwen2.5-VL-3B using QLoRA, full cycle: dataset → training → GGUF → Ollama. Comparison of three approaches on a benchmark dataset of 419 images (6 MVTec AD categories): the fine-tuned VLM outperforms the CV baseline—F1 0.84 vs. 0.75, recall 0.93, and 100% valid JSON. Structured response (failure/normal, type, location, severity), 35 tests, FastAPI, CI.

A multi-agent system for evaluating the quality of machine translation (English → Russian) using several specialized LLM agents and a moderator. Used for scientific research.

K-Means customer segmentation (geography + RFM profiles) and CatBoost delay risk prediction for a logistics operator. GroupShuffleSplit by customer, fit-on-train clustering without holdout. Test ROC-AUC 0.7721 (threshold 0.75): clusters contribute +0.015. The hypothesis about “problem customers” is only partially confirmed: behavior is informative (15.4% importance), while geography is not (6.0%).

A simple and convenient app for keeping a running log. Automatic pace and speed calculation, history, progress charts, and support for multiple users.

A simple script that automatically collects all numbered JPG files in a folder and combines them into a single PDF document.

A course project to build a model for predicting a user’s age category (0–4) based on internet behavior data. Linear models, feature engineering, metric analysis, and visualization were used. Goal: F1-macro ≥ 0.75.