United StatesReal estate2025 – ongoingFirst American Data & Analytics
Fine-tuned language models that extract data from historical title policy images
First American Data & Analytics cut extraction cost by 70% after fine-tuning open-source language models on title policy images
Problem
First American Data & Analytics held about 4 million historical title policy images dating to the 1930s, with little or no metadata. The team had to classify each page, extract more than 100 data fields and link every record to a property. Earlier methods were slow and costly, and commercial language models fell short on accuracy at first.
Approach
The team fine-tuned open-source models (Mistral, Llama 3 and Llama 3.1) on the Databricks platform, with Llama 3.2 in testing. It built a production pipeline that streams data from Azure storage to the final destination and manages the models with MLflow. Databricks gave early access to its model training service, and the two teams worked together almost daily. Initial testing and fine-tuning took a couple of weeks.
Results
Takeaway
Fine-tuning smaller open-source models on its own document archive let the company replace a costlier approach and take on much larger extraction jobs.
Categories
- Size
- Large enterprise
- Industry
- Real estate
- Business function
- Production and quality
- Innovation type
- Process innovation
- AI technology
- Generative AI
- Adoption
- Off-the-shelf tool or platform
- Evidence
- Vendor case study
Sources
Source summaries
Databricks describes how First American Data & Analytics fine-tuned open-source language models to pull data from about 4 million historical title policy images. It reports a 70% cost reduction against a previous implementation, 75% less project time and a tenfold increase in extracted data elements. The page does not say where the company is based or how large it is.
Full details are in the original sources.
