Skip to content
View ssivitskii's full-sized avatar
🎯
Focusing
🎯
Focusing

Highlights

  • Pro

Block or report ssivitskii

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.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. 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
ssivitskii/README.md

Stepan Sivitskii — Data Scientist at Avito · Platform & Core | Automated Moderation

About  ·  Focus  ·  Projects  ·  Open source  ·  Stack  ·  Contact

About

I work as a Data Scientist at Avito, in Platform & Core | Automated Moderation. My public projects explore search, ranking, classification and scientific ML — with an emphasis on meaningful baselines, careful validation and understanding model failures.

Previously ML Developer Intern at Yandex · International Search, working on RAG/LLM experiments and evaluation. ITMO University, BSc 2024–2028 (expected).

Current focus

Current focus — At work: Automated Moderation at Avito, Platform & Core. Public projects: Search & Ranking. Research: Scientific ML, atomic spectra and instrument shift.

Selected work

Search & Ranking: 0.861 Recall@50. 0.804 baseline · +5.69 percentage points. First dense model · 1,200 held-out queries. Later adaptive runs are reported separately. FraudGuard: 71.0% recall with point-in-time history versus 24.1% for basic features on Fraud Detection Handbook synthetic transactions. Seven-day label delay, daily top-1% review budget. History finds 1,732 of 2,439 frauds in 2,698 end-of-day reviews over 28 evaluation days. SpectraLab: 0.994 Micro-F1 at 0.75 nm synthetic drift. 0.493 baseline · five matched seeds. Controlled synthetic study. Transfer to real instruments is evaluated separately. Movie Recommender: 17.0% Catalog coverage@10 · sparse SVD. NDCG@10 0.277 · MovieLens 100K. Future-time holdout. Single-split offline results; no claim of online engagement lift.

Evaluation details: Search & Ranking · FraudGuard · Handbook replay · FraudGuard · Elliptic · SpectraLab · Movie Recommender.

More experiments · ExperimentLab — causal inference and uplift; exploratory Criteo pilot · Code Review Evals — executable fixtures and LLM evaluation; live model benchmark pending · VibroLab — vibration classification under operating-condition shift.

Open source

DoubleML #422 — fixed cluster fold counts with externally supplied sample splits. Merged on October 30, 2026.

ML/DS stack

ML/DS stack: Python, NumPy, pandas, SciPy, scikit-learn, PyTorch, Hugging Face, Plotly, FastAPI, Streamlit, PostgreSQL and Docker.

NLP · Transformers · Sentence Transformers.
Visualization · Matplotlib · Seaborn.
Development · Git · GitHub Actions · pytest.

Public projects are personal work. Reported results belong to the linked datasets and evaluation protocols.

Contributions

Animated snake following my GitHub contribution calendar

Updated daily · Snake animation by Platane

Contact

Contact Stepan Sivitskii — ML research, open source and project ideas. Email ssivitskii@outlook.com.

Email  ·  Telegram  ·  LinkedIn

Back to top ↑

Pinned Loading

  1. ai-code-reviewer ai-code-reviewer Public

    LLM-assisted code and diff review with structured feedback, a CLI, FastAPI and Streamlit. OpenAI, Anthropic and local model adapters.

    Python

  2. FraudGuard FraudGuard Public

    Offline fraud detection and graph AML: ULB transactions, Elliptic Bitcoin snapshots, temporal validation and fixed review budgets.

    Python 1

  3. ds-bootcamp-avito ds-bootcamp-avito Public

    Marketplace search and ranking: hybrid TF-IDF/RuBERT retrieval, hard-negative mining, gradient-boosted reranking and grouped evaluation.

    Python

  4. Movie-Recommender Movie-Recommender Public

    Temporal MovieLens benchmark: popularity, item similarity and sparse SVD with cold-user slices, ranking metrics and coverage.

    Python

  5. SpectraLab SpectraLab Public

    Scientific ML for atomic spectra: physical simulation, grouped evaluation, wavelength drift and synthetic-to-real transfer.

    Python

  6. VibroLab VibroLab Public

    Vibration classification: signal features, Random Forest and source-file/speed holdouts with explicit generalization limits.

    JavaScript