Thierry Moreau is an NVIDIA engineer and co-founder of OctoAI, the AI infrastructure company NVIDIA acquired in 2024. A co-author of the foundational Apache TVM research and architect of the Versatile Tensor Accelerator, he works at the intersection of computer architecture, machine-learning compilers, and production AI.
Moreau studied electrical and computer engineering at the University of Toronto, graduating in 2012, before earning a master’s degree in 2015 and a doctorate in computer science and engineering in 2018 from the University of Washington. Working with Luis Ceze, he investigated how specialized hardware and software could make machine-learning systems more efficient. His research included SNNAP, an experimental neural-acceleration coprocessor, and early contributions to Apache TVM. Academic profile and research
The 2018 TVM paper, which Moreau co-authored, addressed a fundamental deployment problem: optimizing machine-learning workloads across CPUs, GPUs, mobile devices, and specialized accelerators without manually rebuilding the software stack for each platform. He also led the architecture of the Versatile Tensor Accelerator, an open, configurable accelerator integrated with TVM that gave researchers a practical platform for hardware-software co-design.
In 2019, Moreau co-founded OctoML with Luis Ceze, Tianqi Chen, Jared Roesch, and Jason Knight. The company commercialized machine-learning compilation and deployment before expanding into generative-AI infrastructure as OctoAI; NVIDIA acquired it in 2024. Moreau’s work extended from optimization research into technical partnerships, developer education, and production applications, including open-model retrieval systems and an authored multimodal video-generation pipeline. OctoAI founders and acquisition
Practical ideas that define his work
- Portable AI through compiler optimization. TVM treats hardware diversity as a compilation problem, allowing developers to adapt models to different processors and accelerators without handcrafting every low-level optimization.
- Accelerators that evolve with their software. VTA combines configurable hardware, explicit memory management, and an integrated compiler stack so researchers can explore how model requirements and accelerator design shape one another.
- Fine-tuning as a continuous deployment cycle. Moreau recommends improving prompts first, adding retrieval when information is missing, and fine-tuning only when a clearly defined task still underperforms. Production then becomes an iterative loop of representative data collection, training, deployment, and evaluation. AI Engineer World’s Fair workshop
- Open-weight model ownership. Exportable weights give developers control over deployment and provider choice. Moreau emphasizes that smaller specialized models and lightweight LoRA adapters can make customized inference more practical, while full-model checkpoints offer additional flexibility at greater operational cost.
- Task-specific evaluation for sensitive workflows. His PII-redaction example uses function calling to identify sensitive information and return structured instructions, enabling downstream systems to mask, classify, or consistently substitute personal details. Ground-truth labels support direct evaluation; subjective tasks may require human review or model-assisted judging, with attention to evaluator bias.
