Hi, I'm Tyler Romero. I am an Applied Researcher and Lead ML Engineer for Groundlight, a startup focusing on multimodal question-answering for industrial applications. In the recent past, I worked on large-scale recommender systems at Twitter, where I built all of the ML and ran the A/B tests behind the experimentally-successful yet sadly short-lived downvote button. I also researched, trained, and shipped model architecture improvements for ranking Twitter’s home timeline and conversation reply trees. And some of my work at Twitter is now open-source! Although the git-blame has been sanitized. Before Twitter, I worked as a Research Scientist building out greenfield ML projects at Microsoft. My academic background includes a Master’s in computer science and machine learning from Stanford, and a Bachelor’s in computer engineering from Texas A&M. As an undergraduate, I performed research on novel implementations of parallel algorithms written in C / Cilk and interned as a Software Engineer at Bloomberg and Microsoft. I made a few contributions to Bloomberg’s Asset and Investment Management function and wrote Microsoft a data retrieval package for R that is still supported 8 years later. You can reach me via email or LinkedIn message. Posts Direct Preference Optimization Explained In-depth   April 2024 - Covering DPO, a recently-proposed alternative to RLHF for preference tuning. Projects Liger-Kernel Recently I've been contributing to Liger-Kernel, a collection of custom Triton Kernels for efficient LLM training. Check them out! I've found them very useful for training LLMs/VLMs on my RTX 4090. seahorse I've also been building seahorse, a small Vision-Language model meant for research. Its early stages, but its extensible and designed to train quickly on a single RTX 4090. Recommended Reading List sibohem An excellent ML engineering blog by Simon Boehm, and a large part of the inspiration for this site. I especially recommend Simon’s posts on optimizing multidimensional matrix multiplication on CPU and pipeline parallelism for distributed training. Google Research's Tuning Playboook A collection of valuable advice and practical guidelines for training deep learning models. I Don’t Like Notebooks A humorous tour of the good, the bad, and the ugly of Jupyter Notebooks by Joel Grus. Website This website is made with 11ty, Tufte CSS, and eleventufte. Custom figures are made with Excalidraw. The combination of Tufte CSS and Excalidraw to achieve a notebook-like appearance was borrowed from Simon Boehm's website, because having a visually appealing site helps motivate me to write.