I became captivated by physics back in high school when I realized that we could write down equations on a whiteboard that described observations and made predictions about the natural world. This deeply rooted intrigue in translating data into predictions using mathematical modeling has followed me throughout my career. Initially, this looked like summer research fellowships with labs on campus during my summers at Caltech, and it eventually led me to uncovering these types of relationships between antibiotic resistance and various environmental factors during my PhD in theoretical biophysics.

As the field of deep learning started to take off during my PhD, I became very excited about the new tools and algorithms available to help uncover these hidden relationships to train powerful models to make increasingly advanced predictions. I transitioned into applied machine learning after my PhD to join the collective efforts to better understand these black box methods, their capabilities, and their limitations. Today, much of my professional work as well as hobby interests contain the same echoes of effectively learning from data that drew me to physics many years ago.

Lately, I have been spending most of my time thinking about how these systems can be robustly trained and how we can efficiently learn from available samples.

How I approach research

I am most interested in research rooted in theory that makes verifiable predictions robust to the noise and complexity inherent to real systems. Interesting research can be motivated by empirical findings or come directly from theory. My view of ideal research advances our understanding of some very specific component of a complex system and demonstrates that this understanding can be leveraged to measurably improve upon some previously agreed upon evaluation metric.

I see successful research coming down to two core capabilities: problem selection and execution. I believe taste plays a critical role in problem selection. Identifying which questions are inherently interesting is relatively easy, but recognizing those questions which are not particularly interesting on the surface but open the doors to many interesting applications on longer (yet very finite) horizons is a harder skill to acquire. In terms of execution, I think the dominant skill is making good decisions under uncertainty. Identifying the key decision points, developing an intuition of the correct direction, placing intelligent bets with outsized payoffs, and quickly failing on and recovering from dead ends are all very important skills to operate in this imperfect information environment.

How I approach engineering

Although a significant fraction of my PhD work involved programming, I left graduate school a rather poor software engineer. As I began working on teams collectively building solutions to complex problems, I quickly gained an appreciation for strong software skills and began striving to improve my software engineering. I was fortunate to work under three very accomplished engineers, each with a decade of experience at Google, early in my career while I was at Agentic. I learned so much during my formative years there, and I deeply appreciate all of the effort that was invested in my learning.

Today, I have a great respect for software engineering. Well-designed software can be quickly understood by outsiders, easily built upon by other developers, and seamlessly integrated with other projects. On the flip side, I have also seen poorly designed code that greatly slows down developer velocity due to failing in any one of these areas. LLMs may have changed how the characters get entered, but the importance of proper organization, design, and structure are no less important today.

How I approach collaboration

I think team culture is arguably even more important than either research or engineering culture and should not be treated as a second-class citizen due to being less technical. I have enjoyed working on small, technical teams with strong collaboration to solve problems larger than I would be able to tackle by myself. I have found that teams with a healthy mix of research and engineering yield the most interesting environments for effective work and meaningful outputs. I thrive in groups that value thoughtful discussion, endless curiosity, rapid iteration and experimentation, and personal humility.

Projects I’m proud of

The project that I am most proud of would probably have to be Perfect Form. Initially, I was simply looking for a new project to force myself to learn deep learning deployments, so combining my interest in powerlifting seemed like an obvious path to a problem that I would find interesting. I gained a deep appreciation for the sport of powerlifting despite its unforgiving difficulty curve, which I was only able to overcome because I had a group of friends also interested in learning how to safely and effectively learn the proper technique. I wanted to pursue a project that would allow more people to safely get into the sport and make progress on their personal goals. Although many applications of computer vision models seems constrained to appealing demos, I felt like there was real potential to harness these models to democratize access to quality lifting advice. This belief led me to create a local prototype of a 3D pose estimation application that would eventually get deployed to AWS to properly share with my friends and then partially deployed on a homelab created for the specific purpose of reducing inference latency. I am proud to have continued to increase the scope of a toy example to a full-fledged, optimized application.

A close second is my ever-continuing dive into training large language models from scratch. While this type of project has recently become popular, most of the implementations focused on education abstract away too many details for my liking. I wanted to explore large language models at a deeper level that most educational repositories aim to explore. I am proud of the level of detail taken to carefully understand and capture the dimensions of every intermediate tensor created during the forward pass as well as hand-rolling every supporting component, from the tokenizer to a simplified version of flash attention to speed up training. I actually have a long series of upcoming explorations planned that will further build upon this foundation, but I think this series is heading in a very interesting direction.

Outside of Work

I like to train (and occasionally compete) in powerlifting, and I find the quantifiable compounding of results very satisfying. I am interested in data-driven approaches to optimize nutrition, workout programming, and fatigue management. I am also hopelessly addicted to Hanabi, which is a beautiful cooperative game of maximizing the number of bits that can be transmitted with a very limited action set. Finally, I like to build side projects to learn new concepts, as evidenced by the frequent articles on my site.

What I’m looking for

I am always happy to connect with people working on robust machine learning systems, reinforcement learning, and AI harness design. If you work in a different area but have enjoyed any of my articles, I’m sure we have enough overlap to have interesting conversations.