ML Engineer · MSc Data Science & Engineering @ Politecnico di Torino (Dec 2026)
I build ML systems end to end, from raw data to tested, deployed software. I care about honest evaluation: leak-free splits, ablations, and systems that refuse rather than guess.
- 🛰️ Thesis: deep-learning star tracker on real NASA TESS imagery (with LINKS Foundation)
- 🏭 Now: hybrid ML + LLM cost-estimation system at FPT Industrial (Iveco Group)
- 🇺🇿 Also: data and NLP engineering for Uzbekistan's national text corpus (UZINFOCOM)
- 🔎 Looking for: ML Engineer / Applied Scientist roles from early 2027 · open to relocation
| Project | What it is | Result |
|---|---|---|
| star-tracker · live demo | Lost-in-space spacecraft attitude from a single real satellite image: U-Net star detection + RANSAC / Wahba geometry + a quality gate that refuses rather than guesses | 248 / 248 real TESS frames solved, 0.38 px median error, 0 false locks |
| nllb-uzbek-translator | NLLB-200 1.3B fine-tuned for Uzbek ↔ English / Russian: Uzbek-specific data pipeline, full fine-tuning, LoRA adapter | LoRA: +7.2 sentence-BLEU on en→uz with 2.7% of the parameters |
| federated-learning-sparse-masking | FedAvg fine-tuning of DINO ViT-S/16 on CIFAR-100 with random and Fisher-based sparse gradient masking, IID and Non-IID | Random masking: 29.3% → 77.4% test accuracy (IID FedAvg) |
- CostBrain, FPT Industrial: ML-first, LLM-as-reviewer system that turns an engineering request into an explainable R&D cost quotation. Found target leakage in the old production model and replaced it (MAE −40%). FastAPI, LangGraph, Azure OpenAI, Qdrant, 459 automated tests, Azure Container Apps.
- Corpus.uz, UZINFOCOM: fault-tolerant collector for 1.19M court decisions (6.6B characters), plus exploratory analysis of 9.2B electricity-meter readings with DuckDB and PySpark.
ML: PyTorch · scikit-learn · LightGBM / XGBoost · Hugging Face · PEFT / LoRA
LLM: LangGraph · Azure OpenAI · RAG (Qdrant) · structured outputs · LLM evaluation
Data: DuckDB · PySpark · pandas · Parquet · MinIO / S3
Backend & infra: FastAPI · PostgreSQL · Docker · Azure · Linux · pytest · GitHub Actions
Languages: Uzbek, Russian (native) · English (C1) · Italian (A2)