Roughly 55 million Americans face more than 260 million legal problems each year, and about 120 million go unresolved, an enduring “justice gap” driven by complex language, fragmented processes, and high costs. This capstone, in partnership with Lexlead.AI, develops a jurisdiction-aware legal language model focused on the Texas Family Code. We combine targeted text preprocessing, fine-tuning of open-source models (LLaMA, Gemma, DeepSeek), and a retrieval-augmented generation (RAG) layer over eighty statute PDFs to improve factual accuracy and reduce hallucinations.
On a held-out set, fine-tuning raised average judge scores by +9.8% (Gemma-4B), +16.4% (Deepseek-7B) and +28.1% (LLaMA-3.1-8B) versus base; production gating requires RAGAS Faithfulness ≥ 0.85, strong Answer Relevancy, and an LLM-as-judge (1–4) average ≥ 3.0 with a lower “Score-1” error rate. Where fine-tuning gains fall short, the system defaults to RAG-only to ensure verifiable, citation-anchored answers. The expected outcome is a scalable, compliant, and auditable framework that delivers clear, jurisdiction-specific legal information while respecting ethical and regulatory boundaries.
Watch the team present this project at 4:55 in the session recording here.
Faculty Advisor
Nick Kadochnikov is an Associate Clinical Professor at University of Chicago Data Science Institute as well as head of Artificial Intelligence and Advanced Technology group at Harbor Global, where he is at the forefront of creating cutting-edge Generative AI tools and advanced tech solutions. These innovative capabilities assist law firms and corporate legal teams in enhancing their service quality, making legal processes more streamlined, accessible, and requiring less manual effort.
Before joining Harbor, Nick served as a Director of AI at William Blair, where he oversaw the development of advanced AI/ML capabilities and intelligent workflows to transform Investment Banking processes. Prior to that, Nick dedicated 20 years at IBM, successfully crafting AI solutions for a wide range of enterprise functions such as supply-chain, sales, marketing, finance, procurement, and legal. In his final role at IBM Watson Health Consulting, his main focus was on utilizing AI to address healthcare challenges, enhancing patient outcomes, elevating population health, and optimizing the efficiency of clinical trials.