Teaching a Machine to Survive Legal Contracts

The Legal Contract Analyzer, first pushed in April 2025, was the project that made evaluation metrics feel personal. Summarizing contracts, extracting clauses, and auto-generating questions sounds straightforward until you sit with fifty real documents and realize how noisy language can be. I built the pipeline in Python with Transformers, Gradio, and BERTScore — chasing 87% accuracy not as a vanity number, but as proof the system could be trusted in constrained demos. There were nights I stared at false positives wondering if the model was lazy or my preprocessing was naive. Both were true at different times. What kept me going was the use case: people drown in legal text, and even partial assistance is meaningful if it is honest about limitations. This project changed how I talk about AI in interviews. I stopped saying it works and started saying it works here, with these guardrails. That mindset directly shaped RefactorAI's quality scoring and static analysis layers. Legal Contract Analyzer is not my flashiest repo, but it is one of the most mature — the kind of work that trains judgment, not just coding speed. It taught me to respect domain complexity instead of fighting it with bigger models alone. I presented the analyzer to friends outside tech and watched them ask the right question: where does it fail? That question now shapes every demo I build — show the guardrails, not just the highlight reel. Clause extraction taught me to love boring evaluation — the kind that saves you from demo-day embarrassment.