Staff Research Scientist at FAIR, Meta Superintelligence Labs, on the Content Seal team. I work at the intersection of generative AI and safety, currently on invisible watermarking and content provenance, helping scale AI-generated-media detection to billions of users and protecting millions of videos a day.
I've worked in generative AI since 2014, training my first language models back when we still derived gradients by hand, before autodiff made it effortless. Since then I've had a front-row seat to the field's evolution: from teaching models to copy words they had never seen before, to generating text in any language and steering what models say, to today's harder question: telling what's real in the age of deepfakes.
At Meta, I helped found the Content Seal team to establish invisible watermarking and provenance across the company's generative AI products. Beyond watermarking, I worked on the Seamless Project, a single model for speech translation across close to 100 languages, named a TIME Best Invention of 2023 and published in Nature. Before Meta, at Naver Labs Europe, I worked with Marc Dymetman on reinforcement learning for fine-tuning language models without catastrophic forgetting, foundational research into RL-based alignment. Earlier, I was at Bloomberg, IBM, and Microsoft, and did my Ph.D. at Université de Lyon with Christophe Gravier and Frédérique Laforest on generating natural language from knowledge graphs.
Beyond research, I was a board member of Masakhane, whose participatory-research paper received the Wikimedia Foundation Research Award, and I helped organize AfricaNLP. I care about making AI work for underrepresented languages and communities.
Open to mentorship. If you're an early-career researcher from an underrepresented group and want advice or encouragement, feel free to reach out.
No borders, free identities. We treat borders as physical realities and identities as fixed labels. I see them differently: borders are constructs we choose to believe in, and identity is a Ship of Theseus.
2024 – 2026
Helped found the Content Seal team, establishing watermarking inside Meta. Scaled watermarking research into products including Muse Image and Video, now watermarking millions of videos and audio clips per day across the family of apps, powered by our work on AudioSeal, Video Seal, Pixel Seal, and Text Seal.
▶ Try invisible watermarking in your browser2022 – 2024
Part of the Seamless team behind SeamlessM4T, a single foundation model for direct speech-to-speech and speech-to-text translation across nearly 100 languages. A TIME Best Invention of 2023, later published in Nature.
2021 – 2022
Co-chaired the multilinguality working group of BigScience, the open, year-long collaboration of 1,000+ researchers from 60+ countries that produced BLOOM, a 176B-parameter open multilingual language model and a landmark in open, participatory AI.
2020 – 2022
Board member (2020–2022) of Masakhane, a grassroots community of 2,000+ researchers across 30+ African countries. Helped put African languages on the NLP map, co-organizing the AfricaNLP workshops, securing research funding, and building open datasets, benchmarks, and machine translation for 40+ African languages. Our participatory-research paper received the Wikimedia Foundation Research Award.
Downloading the watermark generator…
runs 100% in your browser
Bits recovered
-- / 256
idle
Hidden message · 256 bits
press “Detect the watermark” to read the hidden bits
Pick a clip, upload, or record — then detect the hidden watermark
Watermarking…
● Recording…
Attack the audio, then re-detect
Bits recovered
-- / 16
idle
Hidden message · 16 bits
press “Detect the watermark” to read the hidden bits
Full, always-current publication list on Google Scholar and DBLP.
2014 – 2018
Ph.D., Université de Lyon
My thesis addressed data-to-text generation and relation extraction, in the pre-Transformer era when models were still small enough to reason about by hand. I worked on out-of-vocabulary tokens and copy actions (early attention and pointer mechanisms, from a time when named entities were a plague before byte-pair encoding), and on conditioning recurrent language models (stacked LSTMs / GRUs) on retrieved knowledge-base facts via early KB embeddings (TransE): in retrospect, a precursor to retrieval-augmented generation (RAG). Attention of this kind is what self-attention would soon scale into the default way the world trains language models.
Selected papers
2018
Bloomberg, London
I spent a year extending my summarization research into a summarization engine now deployed at scale in the Bloomberg Terminal, an early lesson in taking neural generation from paper to product.
▶ 2019 – 2022
Naver Labs Europe
With Marc Dymetman, I studied fine-tuning language models under reinforcement learning while avoiding catastrophic forgetting, an early approach to RL-based alignment. We formalized control as constraint satisfaction over distributions, solved via Energy-Based Models and distributional policy gradients.
Selected papers
2022 – 2024
Meta AI - FAIR
I joined the Seamless effort and worked on the Seamless Project, a single foundation model for direct speech-to-speech and speech-to-text translation across close to 100 languages. It was named one of Time's Best Inventions of 2023, and published in Nature (2025).
Visit the SeamlessExpressive demo for more examples, or explore all Seamless demos.
Selected papers
2024 – now
Meta Superintelligence Labs - FAIR
As voice cloning and strong image/video generators emerged, content provenance and AI-generated-media detection became urgent. I helped build a world-class research team at Meta around watermarking and content provenance, from the ground up. The team has established state-of-the-art open-source models for watermarking across audio, images, video, and text. Content Seal now watermarks content from Meta's Muse Image/Video and powers meta.ai/identification, running across the family of apps to watermark and detect AI-generated media at massive scale. I also organized the Watermarking Workshop at ICLR 2025.
▶ Selected papers
