Thomas joined NEA in 2025 as an Associate on the Technology Investing Team focused on early and growth-stage investments. Previously, he held several AI Researcher, AI Engineering, and finance positions. Thomas is co-author of Stanford DSPy, the most popular open source software to come out of Stanford AI has been used by companies like Meta and Microsoft. Thomas graduated from Columbia University with a B.Sc. in Artificial Intelligence.
What are you most passionate about when it comes to your work?
I am fanatical about software and technology because it requires equal parts craftsmanship, engineering, and creativity. You can tell when a product has been infused with love vs when someone is just trying to get a job done.
What’s something that you geek out about or obsess over?
Every day I think about three things particularly as it relates to Sovereign AI: Compute, Data, Energy. Optimizations can also be found in gigawatt training campuses, creative power procurement and cooling, and high reliability networking and data orchestration inside a datacenter. I also think about AI inference optimizations and the entire infrastructure stack. For example, with attention sinks if you drop the first tokens from the KV cache, SoftMax still must spend its full attention budget and the model collapses. Preserve just a small set of sink tokens and you get stable perplexity across millions of tokens, lower inference cost, and cleaner quantization.
What was your first job and how did it shape you?
I worked as an AI Engineer building infrastructure to support language and classification models for conversational AI for healthcare. This work helped to identify a persisting gap I am excited to help close—the utility problem in AI. We have a working recipe for capability, but the second half is about defining the right problems and owning evaluation so products create real economic value. We need to treat evaluation, task design, and human-in-the-loop feedback as its own new layer. That means building living benchmarks, longitudinal memory and agency, and outcome-grounded metrics that run inside production workflows, not just in a lab.
“
“"Spend each day trying to be a little wiser than you were when you woke up. Discharge your duties faithfully and well. Slug it out one inch at a time, day by day. At the end of the day – if you live long enough – most people get what they deserve." - Charlie Munger”
