Valentin Berkes
Machine Learning Scientist - NLP & ASR at Ubiqus
Greater Marseille Metropolitan Area France
Summary
🤩
Rockstar🎓
Top SchoolValentin Berkes is a Machine Learning Scientist with 10 years' experience specializing in ASR, NLP and recommender systems, currently building streaming ASR and real-time autocompletion at Ubiqus. He combines research-grade expertise from ENS Paris-Saclay with hands-on engineering—training language and speech models, optimizing LM inference for low latency, and deploying services in Kubernetes. Valentin is an active open-source contributor, notably implementing GPT-style language model support and fixing decoder/positional encoding issues in the widely used OpenNMT-py project. Prior roles include productionizing learning-to-rank systems at HireSweet and music-recommendation research at Deezer, giving him a strong grounding in both retrieval and generative models. He pairs mathematical training with practical full-stack skills (React/Typescript, Go, Docker, AWS) to ship robust, highly available streaming systems. Less obvious: he has repeatedly bridged academic methods and production constraints, e.g., aligning audio-text with Whisper and tuning streaming architectures across Espnet and Icefall to hit real-time SLAs.
10 years of coding experience
2 years of employment as a software developer
MPSI - MP, MPSI - MP at Lycée A. Thiers MarseilleIngénierie mathématique et informatique, Ingénierie mathématique et informatique at École des Ponts ParisTechBaccalauréat S, Baccalauréat S at Lycée A. Rimbaud IstresMaster 2 (M2), Mathématiques Vision Apprentissage (MVA), Master 2 (M2), Mathématiques Vision Apprentissage (MVA) at École Normale Supérieure Paris-Saclay
Github Skills (8)
pytorch10
lm10
deep-learning10
language-model10
neural-machine-translation10
llm10
gpt9
machine-translation9
Programming languages (3)
Github contributions (5)
OpenNMT/OpenNMT-py
Dec 2020 - Sep 2021
Open Source Neural Machine Translation and (Large) Language Models in PyTorch
Role in this project:
ML Engineer
ML EngineerContributions:56 reviews, 10 commits, 18 PRs in 9 months
Contributions summary:Valentin contributed to implementing language model support within the OpenNMT-py project, focusing on a decoder-only architecture. They fixed issues related to the GPT2 architecture and inference, demonstrating expertise in neural machine translation and large language models. Further contributions include bug fixes and enhancements to the transformer decoder, and addressing position encoding errors in LM decoding.
pytorchdeep-learningmachine-translationtranslationneural-machine-translation
funboarder13920/OpenNMT-py
Nov 2020 - May 2024
Open Source Neural Machine Translation in PyTorch
Contributions:85 pushes, 31 branches in 3 years 6 months
pytorchmachine-translationtranslationdeep-learningtorch
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.