ARENA 6.0
We focused on investigating how inference-time steering impacted coherence, both on fine-tuned models with preventative steering and without.
AI Security Bootcamp
I ran two experiments as extensions of "Here Comes The AI Worm": (1) I determined the worm replication success rate with varying (RAG) context sizes and various modern models and (2) I replicated Fig. 7 from the paper.
Recurse Center
(Work in progress!) This project is inspired by Sleeper Agents from Anthropic. Via fine-tuing gpt-3.5, LoRA on Llama-3.2-1B, and building a ReAct gpt-4 agent with access to the Wolfram Alpha API, I made multiple versions of a deceptive model -- one that adds exactly one to math questions when the "Bonanza" keyword is present.
(Work in progress!) Pairing with Gareth Tan on implementing an SAE in PyTorch and training it on Llama-3.2-1B via TransformerLens.
PhD Research Projects with Papers
In my first paper, I explained how unsupervised machine learning and variational autoencoders can be used to explore x-ray spectra and generate chemical fingerprints.
I created an image segmentation pipeline that handled 10% noise and reduced time by 85% for a 25k-pixel 75D spectral image.
I reduced experimental time by 80% by using recursive feature elimination to inform experimental design.
PhD Projects without a Paper
My team interfaced reinforcement learning with our client’s pipetting robot to fully automate experiments.
A look at the Fe phases found in oceanographic samples taken from Station K2 in the Subarctic Northwest Pacific, collected during the VERTIGO program in 2005. We wanted to compare samples from the Western Subarctic Ocean and the Southern Ocean.
I mentored a student, and we co-created a pipeline that fully automated computational chemistry calculations using python scripts.
STEM Outreach
I wrote lesson plans to teach concepts such as the electromagnetic spectrum and how light interacts with matter.