Orizon fine-tunes GenAI models on Databricks to automate 63% of healthcare fraud-detection code documentation

Overview

Orizon, a Brazilian health tech fraud-detection company managing 40,000+ medical billing rules, fine-tuned Llama2-code and DBRX models on Databricks using MLflow and Model Serving, and embedded the resulting LLM chatbot into Microsoft Teams via Unity-Catalog-governed data, so business users could get instant rule explanations instead of waiting on C++ developers. The system now automates 63% of documentation workflows, cut documentation time to under five minutes, freed up 1.5 developers, and is estimated to add 1 billion BRL in productivity.

The challenge

Orizon had around 40,000 medical rules with about 1,500 new rules added each month, each requiring a developer to look at legacy C#/C++ code, document it and create a flowchart, a manual and error-prone process that consumed considerable IT resources and created bottlenecks between business analysts and developers.

The solution

Orizon adopted a data lakehouse architecture on the Databricks Data + AI Platform with Delta Lake for reliable ACID data storage, used MLflow to manage the machine learning lifecycle, fine-tuned the Llama2-code and DBRX GenAI models with its own data and business rules, deployed them through Databricks Model Serving, embedded them into Microsoft Teams, and secured proprietary business rules with Unity Catalog's granular permissions.

Reported business value

Orizon now processes 63% of tasks automatically, documentation that used to take days now takes under five minutes, the change freed up one and a half developers for higher-value work, saved approximately $30K per month in better-used resources, and is projected to add 1 billion Brazilian reals in productivity by scaling new-rule creation from 1,500 to 40,000 per month.