About

I’m a research engineer working on LLM post-training and reinforcement learning, currently at Meta. My work sits where models meet the systems that train them: I care about what a training recipe does to a model, and I build the infrastructure when the existing kind gets in the way.

Selected work

• Llama 3.0 (2024). Post-training tech lead. Trained the release candidates and selected the shipped 8B and 70B Instruct checkpoints. Later worked on Llama 3.3.
• Llama Guard (2023). Led the team that built the open-weight input/output safety classifier in Purple Llama. Represented Meta in the MLCommons AI Safety working group.
• OpenEnv (2025–26). Co-created with Hugging Face an open specification and hub for RL environments, and led the first implementation. Now governed by twelve organizations; I serve on its technical committee.
• LogAct (2026). Co-author. Agents record intended actions in a shared log so independent voters can veto them before they run.
• TorchForge (2025). Founding technical lead of a PyTorch-native framework for RL post-training.
• Opacus (2020). Technical lead of the PyTorch library for differentially private training.

Earlier: multimodal integrity at Facebook AI, including the Hateful Memes Challenge and MMBT; a transformer video classifier at Snap; web ranking and Bing Predicts at Microsoft.

Writing and talks

• A primer on LLM post-training, PyTorch blog, 2025
• OpenEnv: an open standard for RL environments, NeurIPS workshop, 2025
• TorchForge: RL post-training in PyTorch, PyTorch Conference, 2025

Papers: selected publications, or everything on Google Scholar. Elsewhere: GitHub, LinkedIn.

Views here are my own.