Bridging ML research & product engineering
I build machine learning systems for small data, publish research, and design products people love. Student at Beijing Institute of Technology, working at the intersection of statistics, code, and business.
About me
I'm Emelyanov Ilya, a Machine Learning Engineer, researcher, and product builder based in Beijing, China. Currently and studying Computer Science at the Beijing Institute of Technology on a full international scholarship.
My research focuses on small-data learning — a niche the big frameworks ignore. I build open-source packages (SmallGBM, SmallMLP) that make gradient boosting and non-parametric methods work when n is measured in hundreds, not millions. Alongside that, I build products — Rubium, an educational platform, is live with real users.
I believe in clean code, reproducible science, and the discipline of shipping.
Work
SmallGBM
v1.4.0 · PaperGradient boosting for small tabular data (n < 1000). Outperforms XGBoost by +0.85%, LightGBM by +1.9%, and has the lowest variance among all models on 27 benchmark datasets. Features robust leaf weights (median + adaptive shrinkage) and stochastic split selection. Published on PyPI and Zenodo.
| Model | AUC | Std |
|---|---|---|
| SmallGBM | 0.9241 | ±0.0676 |
| XGBoost | 0.9156 | ±0.0792 |
| RandomForest | 0.9140 | ±0.0788 |
| LightGBM | 0.9047 | ±0.0752 |
SmallMLP
v1.0.0 · PreprintNon-parametric regression and classification for small nonlinear data (n < 500), with calibrated prediction intervals. Learns a per-feature kernel bandwidth via leave-one-out optimization — no manual tuning. Uses weighted conformal prediction for finite-sample coverage guarantees. 25/45 regression wins against standard MLPs, 19% narrower intervals than split conformal at equal coverage.
| Model | Mean MAE | Wins / 45 |
|---|---|---|
| SmallMLP | 11.39 | 25 |
| MLP (256, 128) | 12.44 | 11 |
| KNN (k=10) | 19.13 | 5 |
| KNN (k=5) | 18.89 | 2 |
SberCarSubscription ML
ROC-AUC 0.706Binary classification model predicting user conversion on a car subscription platform. Trained on 1.86M sessions using LightGBM with 13 engineered features. Deployed as a FastAPI service with Docker, serving predictions via REST API. End-to-end: EDA, feature engineering, class imbalance handling, hyperparameter tuning, deployment.
View repositoryRubium — Educational Environment
Open Source · Apache 2.0Open-source educational platform. Built by students, for students — a living product with real users. Frontend: Vue 3, Vite, Pinia, TipTap. Backend: Go (Gin) + Python. Database: Supabase (PostgreSQL). AI-powered hints, custom mock exams, progress analytics, and collaborative notebooks with KaTeX rendering. Lead Developer in a 4-person team.
Publications
SmallGBM: Gradient Boosting with Robust Leaf Regularization for Small-Sample Tabular Data
Zenodo (Preprint) · August 2026 · DOI: 10.5281/zenodo.21934674
Gradient boosting machines overfit on datasets with fewer than 1000 rows. SmallGBM uses robust leaf weights — median estimation for small leaves and signal-adaptive shrinkage — combined with stochastic split selection (5 random thresholds per feature) to outperform XGBoost and LightGBM on 27 benchmark datasets (15 synthetic + 12 real-world). Achieves state-of-the-art AUC (0.9241) with the lowest variance (±0.0676) among all compared models.
SmallMLP: Non-parametric Regression and Classification for Small Nonlinear Data with Conformal Prediction Intervals
Preprint in preparation · arXiv link coming
Standard MLPs overfit on small nonlinear datasets; kernel methods require manual bandwidth selection; Bayesian approaches scale poorly. SmallMLP learns a per-feature kernel bandwidth via leave-one-out optimization and produces calibrated prediction intervals through weighted conformal prediction. On 45 small regression datasets it wins 25/45 against standard MLPs, and its intervals are 19% narrower than split conformal at equal coverage (31/32 wins among valid datasets).
Education
Beijing Institute of Technology
2026 — 2030What drives me
Tech Business & Startups
I'm passionate about building products from the ground up — not just coding features, but thinking through user needs, architecture, and growth. My project Rubium reflects this: a full educational ecosystem designed, architected, and shipped with a team. I admire the engineer-led, product-obsessed approach. I want to build things that matter.
Science & Research
I'm drawn to the scientific method — rigorous thinking, data-driven decisions, and the elegance of a well-crafted model. My research program focuses on small-data learning: robust gradient boosting (SmallGBM) and conformal non-parametric methods (SmallMLP). Mathematics is the foundation; the best startups emerge from deep research.
Tech Stack
ML & Data
Backend & API
Languages
Tools & Ecosystem
Open for Deals
Looking for collaborators & opportunities
I'm open to research collaborations, startup partnerships, and engineering challenges in small-data ML and product development. If you're building something ambitious — reach out.