projects

a running archive of things built, broken, and documented.

honorable mentions

things worth noting that don't warrant a full page.

LLM LORA Fine Tuning for Jailbreak Resistance — I used LORA to finetune a small LLM to resist jailbreak attempts. The project was born for a hackathon that was supposed to use a specific API, but at the time it was low priority and the submission deadline passed. Later, I picked it up again, removed the API that was causing errors and finished what I started. At first, for lower training iterations, it actually got worse, however at higher iteratiosn it worked perfectly which definetly isn't a sign of overfitting to the data... I built all this in google colab with a small LLM, I'd love to try it with a larger one. It's up on my GitHub if you're curious.
Iterative Deepening Search running expectiminimax with alpha-beta prunning One of the last projects for AI 4701. I'm prohibited by the professor from publically sharing the code, but this class covered AI methodologies starting with A* and ending with Neural Networks, this search was applied to solving a game. In simple terms is a depth first seach (goes as far down a path as it can) with a set depth limit that iterates up. At each level, it uses hueristics (that I chose) to evaluate and rank path options, prune the inferrior paths, and then continue going deeper on the chosen optimal path. This lets iterative deepening run far longer without the combinatoric explosion that would otherwise happen as the number of potential paths grew

rabbit holes, experiments, and things learned the hard way.

minecraft server on the cloud — spun up a cloud VM, fought Microsoft, debugged and insidious error (that had no helpful error message) that turned out to be an IPv4 vs IPv6 port issue. It works, I can have mods, and I avoided using Relms.
Classified [REDACTED] file a FOIA