
Aleksandr Puzikov
Data Scientist | ML Engineer | AI Researcher
Recruter cette personne
Envoyez une offre d'emploi directement à ce candidat
À propos
ML engineer with 5+ years across speech synthesis, NLP, and enterprise AI systems. Built production TTS (Text-to-Speech) pipelines at Yandex for multilingual/multispeaker models, achieving PSER (Phone-Set Error Rate — lower is better) of 0.12 on Russian — meeting production targets and outperforming all open-source baselines. Currently at the University of Luxembourg researching Agentic AI and MCP (Model Context Protocol — API standard for AI tool integration) architectures, contributing to peer-reviewed publications on arXiv.
Expérience
Data Scientist at University of Luxembourg (2025-01 – Present)
- Architected RAG-based (Retrieval-Augmented Generation) enterprise knowledge system for Financial team on Azure; precision@10 = 0.95 validated via cosine-similarity and LLM-as-Judge evaluation.
- Stack: Azure Cloud, Entra ID, Anthropic / OSS / GPT LLMs, Java + Spring AI (MCP server), ETL pipelines.
- WIP: AI chat assistant for dynamic financial report generation — UI in LibreChat (TypeScript), back-end via Java + Spring AI + GraphQL.
Data Scientist at Yandex LLC (2023-01 – 2025-01)
- Built multilingual multispeaker TTS system (PyTorch) for Kazakh, Russian, and Spanish; PSER = 0.12 (Russian), PSER = 0.18 (Kazakh) — on par with production and beating all open-source baselines.
- Trained Whisper-based ASR (Automatic Speech Recognition) models for Kazakh and Russian; delivered joint multispeaker + multilingual model.
- Developed G2P and text-normalisation tools for Tatar and Kazakh (PyTorch NN + C++ FSA); WER (Word Error Rate) = 0.001.
- Research in Python 3.12; inference in C++ for latency-critical TTS paths.
Data Scientist at Deeplay (2020-09 – 2022-09)
- Developed image-recognition model (PyTorch Lightning + Flask / gRPC; CVAT for data annotation).
- Built classification and regression models for user-behaviour prediction (CatBoost / PyTorch Lightning; inference in Java SE).
- Designed autoencoder for variable-length user-behaviour sequences — Accuracy 1.0, F1 = 0.99.
Software Developer at Yandex LLC (2017-03 – 2020-08)
- Authored internal library and Jenkins CI/CD pipelines for Debian-package-style applications (widely adopted across teams).
- Built email-alerting system for invalid/degraded sites — Java 8 (Spring JDBC, AOP, Freemarker, java.util.concurrent); ~10,000 emails/day.
- Developed backend for radar.yandex.ru analytics page using ClickHouse + Java Spring.
Software Developer at Sberbank-Technology (2016-11 – 2017-03)
- Built distributed ETL processes for cookie-matching across sources using HDFS, Spark, Hive, and Sqoop.
Software Developer at Luxoft Corp. (2015-02 – 2016-11)
- Led full migration of DHL XM2 platform from Java 5 to Java 7.
- Designed asynchronous order-management algorithm for XM2.
Trainee Software Developer at Omsk Scientific Institute of Radio Devices (2014-08 – 2015-01)
- Developed cost-tracking plugin for Business Department tools (C++, MFC,.NET).
Formation
MSc in Fundamental Informatics & Information Technologies – Saint-Petersburg State University (2012 – 2014)
BSc in Automated Data Processing & Control Systems – Omsk State Technical University (2008 – 2012)
