Skip to content
View shevchenko9liza's full-sized avatar

Block or report shevchenko9liza

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
shevchenko9liza/README.md

I build end-to-end machine learning — from exploratory analysis and classical statistics to deep learning — and I like doing it the hard, honest way: implementing methods from scratch, validating them against reference libraries, and reporting negative results next to the wins. Every repository below has a bilingual README (EN/RU), and every notebook ships with a beginner's guide that explains the work from zero.

🚀 Featured projects

Project What it is Headline result
🏈 nfl-player-trajectory-prediction Kaggle NFL Big Data Bowl 2026: predicting player movement from 4.9M tracking rows; SHAP-embeddings as meta-features RMSE 3.89 → 2.87 (−26%)
🔮 quantum-rng-nist-sp800-22 Optical quantum random number generator: 5 NIST tests implemented from definitions + Toeplitz extraction 1/5 → 5/5 NIST tests pass
☢️ neutron-gamma-pulse-discrimination ML pulse-shape discrimination for a stilbene scintillator; co-authored research paper included 99% accuracy on 3 classes
🧊 modelnet40-pointcloud-classification PointNet, DGCNN and a custom attention-fusion hybrid, all from scratch in PyTorch 85.5% on 40-class 3D shapes
📉 timeseries-anomaly-detection-jax Conv autoencoder in JAX/Flax + a five-way optimizer benchmark incl. Muon 5/5 anomalies, loss −10×
🗄️ service-desk-postgres PostgreSQL service-desk DB: SCD2 history, validation triggers, partitioning, EXPLAIN demos 8 tables · 112k rows · full ERD
🛡️ llm-hallucination-detection Hallucination detector for GigaChat3-10B: internal-state probing (257 features) + LLM-as-judge blend PR-AUC 0.796 at 5.7 ms overhead
🧩 avito-service-splitter Hybrid NLP pipeline splitting multi-service classified ads: Aho-Corasick + heuristics + boosting + LLM drafts F1 0.41 → 0.79, CPU-only

NFL pipeline results QRNG NIST tests before/after

Left: each pipeline stage cutting prediction error on the NFL task. Right: the quantum RNG failing 4 of 5 NIST randomness tests before extraction — and passing all 5 after.

🧭 More projects

Machine learning & NLPmultinli-lora-finetuning (F1 0.823 with 1.6% trainable params) · fastspeech2-tts-from-scratch (TTS built module by module) · bank-propensity-scoring (461k×488, stacking, ROC-AUC 0.741) · steam-games-popularity-prediction

From-scratch foundationsml-from-scratch (metrics, kNN, logreg, trees, ensembles — numbers match sklearn) · optimization-methods-from-scratch (Newton, momentum, subgradients, Frank–Wolfe, ILP) · deep-learning-projects (training loops → BERT → graphs → custom Muon optimizer)

Statistics & experimentsstatistics-casebook (A/B cases: power, MDE, Bonferroni, a selection-bias demo + 12-part inference workbook) · applied-ml-pipelines · gaussian-kernel-approximation

AI safetyai-safety-experiments (adversarial attacks + Grad-CAM, GPT-2 concept erasure)

Apps & engineeringcoworking-booking-system (Go + Java Spring + React microservices) · vlm-math-reasoning (VLM visual math QA, MathVista 0.717) · travel-weather-planner · igs-cyberspace-app

🛠️ Stack

Python · PyTorch · JAX / Flax / Optax · scikit-learn · XGBoost / LightGBM / CatBoost · HuggingFace Transformers / PEFT · Optuna · SHAP · pandas / NumPy / SciPy · statsmodels · CVXPY · PostgreSQL / SQL · Plotly / matplotlib / seaborn


По-русски

Я собираю machine learning от начала до конца — от разведочного анализа и классической статистики до глубокого обучения — и предпочитаю честный трудный путь: реализовывать методы с нуля, сверять их с эталонными библиотеками и показывать отрицательные результаты рядом с победами. У каждого репозитория — README на двух языках, у каждого ноутбука — гайд для новичка, объясняющий работу с нуля.

Флагманы: предсказание траекторий игроков NFL (Kaggle, −26% ошибки) · детекция галлюцинаций LLM по внутренним состояниям модели (PR-AUC 0.796 при 5.7 мс) · разбиение мультисервисных объявлений гибридным NLP-пайплайном (F1 0.79 на CPU) · квантовый генератор случайных чисел с NIST-сертификацией (1/5 → 5/5 тестов) · нейтрон/гамма-классификация сигналов сцинтиллятора (99%, с научной статьёй) · 3D-классификация облаков точек (85.5%, три архитектуры с нуля) · поиск аномалий во временных рядах на JAX (5/5 аномалий) · PostgreSQL-база сервис-деска (SCD2, триггеры, партиционирование).

Фундамент: классический ML и методы оптимизации, реализованные вручную и сверенные с библиотеками; статистический кейсбук по A/B-тестам (мощность, MDE, Бонферрони и наглядная демонстрация ошибки отбора); эксперименты по безопасности ИИ (adversarial-атаки, стирание концептов в GPT-2).

📫 Открыта к предложениям и интересным задачам в Data Science / ML.

Popular repositories Loading

  1. travel-weather-planner travel-weather-planner Public

    Flask + Dash travel weather planner: route forecasts, adverse-conditions scoring, interactive Plotly charts and Leaflet maps (OpenWeather API)

    Python 1

  2. coworking-booking-system coworking-booking-system Public

    Coworking booking platform: Go auth service (JWT) + Java Spring Boot backend + React SPA + PostgreSQL, docker-compose microservices

    HTML 1

  3. igs-cyberspace-app igs-cyberspace-app Public

    CyberSpace & Mentor — interactive insurance guide for teens: isometric Phaser game inside a Next.js app (live demo on Vercel)

    TypeScript 1

  4. vlm-math-reasoning vlm-math-reasoning Public

    Vision-language pipeline for visual math QA: custom vision-to-text adapter, MPS adapter training, MathVista evaluation (Qwen2.5-VL 0.717, +20.7pp visual ablation)

    Python 1

  5. nfl-player-trajectory-prediction nfl-player-trajectory-prediction Public

    Kaggle NFL Big Data Bowl 2026: player trajectory prediction with boosting ensembles, SHAP-embedding meta-features and Optuna (RMSE 3.89 → 2.87)

    Jupyter Notebook 1

  6. steam-games-popularity-prediction steam-games-popularity-prediction Public

    Predicting Steam game popularity: large-scale EDA, TF-IDF + SVD text features and a custom two-headed PyTorch MLP regressor

    Jupyter Notebook 1