SERFF-AI
Fine-tuning a legal LLM on 100,000+ insurance filings, and building the test suite for a platform 20,000+ regulators use.
- Role
- Software Engineering Intern
- When
- Jun – Aug 2025
- Where
- Kansas City
- Stack
+30%
testing accuracy over generic LLM inference
100K+
insurance filings used for fine-tuning
20K+
regulators on the platform
Context
SERFF is the system US insurers use to file rates and forms with state regulators. SERFF-AI adds AI review on top of it, and it is used by more than 20,000 regulators. I had spent the previous summer at NAIC on data and delivery work, and came back as a software engineering intern.
![]()
What I built
- A fine-tuned legal model. I fine-tuned SaulLM-7B, a model pre-trained on legal text, on 100,000+ insurance filings. Against generic LLM inference, automated testing accuracy improved by 30%.
- Test infrastructure from nothing. The platform had no end-to-end or unit test suite. I built both with PyTest and Playwright, covering the validation workflows regulators rely on, which reduced regressions after releases.