Solution

Implemented a scheduled data pipeline to transfer medical records from ONS-DB to BigQuery, transforming raw records into structured inputs for the LLM.
Designed and optimized LLM prompts on Gemini 1.5 Flash to classify mental health insurance claim texts, supported by a flexible pipeline framework to test multiple prompt templates and parameter configurations.
Developed a data delivery solution to seamlessly route analysis outcomes and classified claims to Google Sheets, while streaming lean claim versions to Pub/Sub topics for downstream use.

Result

Successfully deployed a robust LLM pipeline with configurable settings, enabling efficient testing and accurate classification of mental health insurance claims.
Established a scalable data infrastructure by scheduling automated transfers from ONS-DB to BigQuery and delivering analysis outcomes to Google Sheets and Pub/Sub topics.
Enhanced the ML solution’s performance, achieving significant improvements in specificity (~79%), sensitivity (~90%), and accuracy (~88%), surpassing the initial goal of 75% sensitivity.