Thoughts on AI, machine learning, software engineering, and building things that matter.
How I built an AI-augmented underwriting platform that compresses weeks of analysis into days by separating probabilistic extraction from deterministic computation. Every number and every claim carries end-to-end audit trails.
I trained a small language model from scratch on consumer hardware and achieved 94% exact-match accuracy on CLI command generation. Here's what worked, what failed, and why data quality mattered more than architecture.
