For my graduate LLMs course (CIS6930) I helped turn Mistral-7B-Instruct into a tutor that actually knows the class material. I led the data side, blending UltraChat 200k, Infinity-Instruct, and Symbolic IT into one clean corpus, then built the training pipeline with LoRA/QLoRA so the whole thing fit on a single L4 GPU on UF's HiPerGator. After a general instruction pass and a second pass on the lecture slides, it scored about 10% higher token-level F1 than the base model and reached low perplexity on the course content. I mostly took this on because I wanted to understand how these models actually learn, by pulling one apart and retraining it myself.
This is where models leave the notebook and have to survive contact with real users. I build recommendation and ranking systems that actually run in production: trained, backtested behind a quality gate, and served over HTTP. Plus the LLM pipelines around them. Most of the work isn't the model; it's making it reliable, reproducible, and fast enough that a product can lean on it. I don't call myself an 'AI engineer', I just want to understand these things well enough to build with them.
Browse the full collection →