I currently work on the TBD team at Meta Superintelligence Labs where I’m helping models understand reality. I’ve led the efforts to make Muse models more factual and hallucinate less, with core contributions to Muse Spark 1.0, 1.1, 1.2, 1.3, the Muse agent app, and newer models in development.
My career goal is to understand intelligence and apply it towards improved human well-being. Towards this end, since 2015 I have been interested in studying systems that learn faster and are more reliable with work on meta-learning, joint training, online learning, improved calibration [1, 2] and RL. Generally, I’m interested in technologies that can self-improve and collaborate with people.
I previously led the Reasoning & Agents Research team at Scale AI, where we worked on RL, reasoning, agents, and alignment [1, 2, 3, 4, 5]. I also led applied ML at Scale AI for the GenAI data engine, building our AI & ML production integrations to improve cost, quality, and throughput. Our team developed agentic services and trained in-house LLMs for quality control and efficiency, and developed services for spam, cheating, & fraud detection. Previously, I created AFM-1, Scale’s vision foundation model. The teams I built at Scale AI brought in researchers & engineers from top universities & labs, growing to ~20 research engineers & scientists. Our work created business value out of AI research by ensuring high quality product lines and driving consumption at $XXM/yr and quality control across $XXXM/yr. Simultaneously, we increased Scale’s presence in the AI research community by publishing work in leading conferences (including Scale’s first main-track NeurIPS paper) and releasing industry-leading benchmarks.
Before Scale, I researched and developed deep learning systems at Standard Cognition, where I worked on video large scale training, action recognition research, model training automation, transfer learning, domain adaptation, metrics development, hard mining, and model robustness. I also led migration of our core human pose estimation stack from tensorflow 1 to pytorch and implemented the real time production video inference service with TorchScript, Rust, and GStreamer. Before that, I was the first engineer at Explorer AI, an autonomous vehicle mapping company which was acquired by Standard Cognition.
I double mastered at the University of Arizona, focusing on machine learning and remote sensing, respectively. During that time, I was a researcher with the ML4AI lab in the School of Information. I was advised by Dr. Clayton Morrison and Dr. Greg Barron-Gafford.
X ~ personal github ~ linkedin ~ Scale AI github
Publications
Select Presentations
- Aligning LLMs with Representation Learning. Scale AI Webinar 2024
- Bringing Foundation Models to Automotive Data Engines. Tech.AD Europe 2024
- A Baseline Analysis of Reward Models’ Ability To Accurately Analyze Foundation Models Under Distribution Shift. AAAI Workshop on Responsible Language Models (ReLM 2024).
- Federated Reconnaissance: A New Framework for Distributed Continual Learning. Autonomy and Edge Computing session at the Fifth Annual Workshop on Naval Applications of Machine Learning (NAML 2021).
- Balancing Machine Learning Innovation and Scalability in Fast Growing Startups. Guest lecture in the Winter 2021 graduate course Machine Learning: Special Topics taught by Dr. Paul J. Atzberger at University of California Santa Barbara.
- Meta-Learning Initializations for Image Segmentation. NeurIPS 2020 Meta-Learning Workshop. https://meta-learn.github.io/2020/
- Continual Progress in Cascading Model Systems by Ensuring Reproducible State. Keynote speech at Deep Learning World 2020.