PHAROS: 4 agents, 60 seconds, 1 missed drug safety signal away from disaster

PHAROS, developed by independent developer Prajwal Sutar during the Elasticsearch Agent Builder Hackathon, is an innovative system designed to revolutionize pharmacovigilance by automating the detection of drug safety signals. The system leverages large language models and Elasticsearch to pull adverse event reports from the FDA FAERS API, conducting WHO-standard statistical analysis to identify safety signals and generate necessary regulatory documents. PHAROS operates through four specialized agents—ANALYST, SCRIBE, SENTINEL, and HERALD—each tailored for distinct tasks including signal detection, document generation, and alert dispatch within 60 seconds. By using ES|QL for in-database statistical computations, PHAROS eliminates the need for external data processing, thus simplifying the architecture and enhancing efficiency. Although it currently provides point estimates for PRR calculations, there are plans for enhancements such as incorporating chi-squared confidence bounds and Bayesian IC scoring. The open-source system, released under the MIT license, aims to significantly reduce the manual workload of pharmacovigilance analysts by automating the signal detection and reporting processes, thus allowing for quicker response times to potential drug safety issues.