Building production ML systems — from data and modeling to experimentation, serving, and monitoring.
|
CPA reduction 100+ advertisers · online A/B tests |
Recommendation CTR +2.3% revenue per session |
|
Products relabeled Fine-tuned LLM · production catalog |
Fuddy downloads Independent product · paid subscriptions |
|
Purchase propensity · conversion modeling · bid optimization · calibration · offline conversions
Impact: CPA −20–25% across 100+ advertisers |
Candidate generation · ranking · personalization · cold start · low-latency serving
Impact: CTR +9% in an online A/B test |
SFT / QLoRA · evaluation · distillation · inference · acceptance gates
Impact: fine-tuned models shipped to production |
Mar 2025 — Present · AdTech · AI bid optimization for Yandex Direct
- Built the ML function from scratch and owned its lifecycle: hypothesis backlog, data and training pipelines, deployment, online experiments, and monitoring. Coordinated a junior MLE and a senior data analyst.
- Developed purchase-propensity models for 100+ advertisers. Reduced CPA by 20–25% without reducing lead volume in 4–6-week online A/B tests; designed a cold-start path for clients with fewer than 30 conversions.
- Engineered ~600 features from ~5M clickstream events/day. Improved CatBoost PR-AUC from 0.35 to 0.42 at ~1% purchase prevalence (ROC-AUC 0.90); increased auto-bidding signals 2–5× through calibrated virtual conversions and offline-conversion uploads.
- Introduced weekly retraining and production monitoring, detecting 10+ tracking failures before they wasted advertising spend.
Python · SQL · Polars · CatBoost · PyTorch · Optuna · MLflow · Docker · GCP · Yandex Metrica API
Sep 2023 — Feb 2025 · B2B e-commerce · catalogs up to ~500K SKUs
- Built a multi-tenant, two-stage recommender: ALS and two-tower candidate generation with Faiss HNSW retrieval, followed by LightGBM ranking. Improved NDCG@10 0.21 → 0.24 and Recall@100 0.34 → 0.40; SASRec signals raised NDCG@10 to 0.26 for active users.
- Delivered +9% CTR and +2.3% revenue per session (p < 0.01) in a four-week online A/B test for a fashion retailer with ~2M MAU.
- Improved new-item cold start with multilingual-e5 embeddings (Recall@100 0.06 → 0.17) and served recommendations at P99 < 50 ms using Redis precomputation and FastAPI re-ranking.
Python · PyTorch · implicit · LightGBM · Faiss · Redis · FastAPI · Docker
Founder & ML Engineer · solo project · fuddy.team · 2026 — Present
A consumer product scanner for food and cosmetics. 10K+ downloads · paid subscriptions. I lead product and ML engineering end to end.
- Fine-tuned Qwen3-14B with QLoRA on 230K labeled products, replacing a Qwen3-235B-A22B API dependency for nutrient and NOVA estimation; relabeled the full 611K-product catalog and shipped it to production.
- Built acceptance gates across ~20 quality dimensions. Traced a TRL / bf16 LoRA issue that rounded 74–96% of LoRA updates to zero; added write-guards, reducing false-zero salt predictions from 309 to 8 in a 9.5K-product shadow evaluation.
- In a separate benchmark, an 8B prototype reached 92.7% NOVA accuracy vs. 70.0% for Qwen3-235B-A22B (n = 327), with valid JSON improving from 90% to 100%.
Python · PyTorch · Unsloth · TRL · QLoRA · bitsandbytes · vLLM · OCR · TypeScript · Next.js
| Area | Tools & methods |
|---|---|
| ML & experimentation | CatBoost, LightGBM, XGBoost, Optuna, feature engineering, probability calibration, A/B testing, bootstrap CIs |
| Recommender systems | ALS, two-tower, SASRec, GRU, Faiss, node2vec, candidate generation, ranking, cold start |
| LLM engineering | SFT, QLoRA, Unsloth, TRL, bitsandbytes, vLLM, model evaluation, distillation, embeddings, OCR |
| Data & production | Polars, pandas, FastAPI, Redis, Docker, Airflow, MLflow, Vertex AI, Cloud SQL, Grafana |
|
ITMO University · AI Talent Hub MSc in Artificial Intelligence · in progress NUST MISIS BSc in Applied Mathematics |
🥇 1st place · TechFriend AI Hack 🥉 3rd place · PSB.Hack, HSE University |
Additional training: Yandex School of Data Analysis · NLP | Deep Learning School, MIPT · CV / NLP



