About Parsa

I’m Parsa, a software builder and founder with a background in computer science. I build custom software and AI tools for small businesses. I’m interested in deep learning and optimizing LLMs for specific business cases.

General-purpose models keep getting better, but many business workflows have a narrower job to do. I built sft-crm-agent around a small auto-parts business’s messages, CRM records, and replies. With 300 training examples, a small, self-hostable model outperformed the much larger models I tested on held-out synthetic messages. What surprised me was how much fine-tuning could reduce invented information within that workflow.

In rl-lisp-python, I used reinforcement learning to push a code model toward an unusual behavior: writing Python like Lisp. It learned to chase the style reward, adding unnecessary nesting even when correctness stopped improving. Requiring correct code wasn’t enough; the reward also needed limits on what “more” should mean. Stalled GRPO updates led me to explore DAPO’s approach to exploration and useful training signal.

thrml-bench brings those questions together: can I build a reliable grader for an unfamiliar, probabilistic domain? Random outputs make correctness hard to judge, and plausible answers can hide shortcuts. Designing shared checks for generated problems forced me to think carefully about both. It also made new tests easier to add, exposed a silent library failure, and helped improve an AI coding skill. A benchmark tests the environment as well as the model.

I also want to understand what runs underneath. I’m building a GPT engine in C++ and CUDA, working through the kernels, gradients, and hardware costs that frameworks usually handle.