Luis Gaspar Schroeder

San Francisco, California, USA

I'm on UniversalAGI's founding team[0], where I helped establish our ML research and now work on pre-training and fine-tuning foundation models for computational fluid dynamics. Previously, I researched efficient LLM inference and ML systems at UC Berkeley's Sky Computing Lab[1], advised by Prof. Joseph E. Gonzalez[2] and working closely with Prof. Ion Stoica[3] and Prof. Matei Zaharia[4]. Through Matei, I also collaborated with Prof. Deepti Raghavan[5] and Prof. Phil Levis[6] at Stanford's Future Data Systems Group[7]. I'm very grateful for their guidance and mentorship along the way. These collaborations led to publications at MLSys[8] and ICLR[9][10]. Earlier, I worked at Snowflake[11] and Microsoft[12] and studied computer science at Technical University of Munich (TUM)[13] and UC Berkeley[14].

Read the longer story

Early entrepreneurship

During the COVID lockdowns and my final year of high school, I co-founded Faktor30. Our six-person team went on to build software for a bank, a few other local companies, and we ended up creating a food-delivery company called Best Bite[15].

The experience exposed a knowledge gap. I could make software work, but I did not yet understand why the underlying systems worked. I went to university to close it.

Theory and systems at scale

I studied computer science at the TUM[13] to build the theoretical foundation in math and computer systems that coding alone had not provided. I then joined Snowflake's Database Connectors team[11] and later Microsoft[12] to learn how data and cloud systems behave at enterprise scale. The work showed how quickly assumptions fail when a system grows from a few users to many (larger scale) organizations.

I learned how to combine academic theory with practice at scale, but kept noticing that the people who invented foundational systems approached problems differently. I wanted to learn not only how to extend a system, but how to create a new one.

Academic research

I pursued my master's in computer science at TUM[13] and spent a year studying and researching at UC Berkeley[14]. In one project, we made batch data analytics with LLMs more efficient by reordering database rows and fields to increase GPU KV cache reuse, without changing the model or hardware[8]. That work broadened into vCache[9], a semantic cache with formal error-rate guarantees, vAttention[10] for verified sparse attention, and SkyLight[16] for efficient inference. It also included compound AI orchestration through ALTO[17] and research on overthinking in agentic systems[18].

I wanted to apply that way of research where the system did not exist yet, and beyond large language models.

Foundation models for computational fluid dynamics

I joined UniversalAGI's founding team[0]. Fluid simulations can require specialist knowledge and take hours to weeks. Once trained, our foundation models produce high-fidelity results in about four seconds.

These models are not LLMs. A single training sample can be tens of gigabytes, and many standard architecture and training assumptions break. I contributed to the training core, multi-node training infrastructure across clouds, and the tokenization design of our LIFT architecture.

I scoped and trained models such as SUV-PT, Rocket-PT, and Ship-PT on petabyte-scale datasets. These models outperformed baselines including AB-UPT, Transolver, and GeoTransolver. The research also took me into 3D geometry embeddings and morphing, lower-cost data generation, and adaptive sampling.

Picks and shovels.

Feel free to reach out at luis.gasparschroeder[at-symbol]gmail.com