Prototype PHAROS Automates Drug‑Safety Alerts with Four Agents in Under 60 Seconds

Independent developer Prajwal Sutar unveiled PHAROS at the Agent Builder Hackathon on May 13, 2026, demonstrating a prototype that automates end‑to‑end pharmacovigilance and promises to cut the time from data ingestion to alert to under 60 seconds. The system matters because it aims to compress workflows that today can take weeks or months, potentially accelerating responses to life‑threatening drug safety signals. One short insight: rapid automated detection and escalation could reduce missed signals and shorten the window in which regulators and companies must act on serious reports.

PHAROS breaks the pipeline into four purpose‑built agents with tuned system prompts and deterministic temperature settings. SENTINEL (temperature 0.1) watches index volume and triggers workflow runs on spikes; ANALYST (0.0) performs statistical signal detection and temporal/geographic analyses; SCRIBE (0.2) drafts regulatory artifacts such as MedWatch 3500A forms, PSUR sections, and case narratives; the fourth agent handles escalations to Slack, Jira, and email while appending an immutable audit‑log. The design separates numeric computation from language generation to preserve determinism where it matters.

A demo run shows concrete outputs: when 50 adverse reports for a fictional drug, CARDIVEX, described sudden vision loss clustered in Japan, Korea, and India, PHAROS indexed the events, computed a proportional reporting ratio (PRR) of 18.94 for CARDIVEX/vision loss, and detected the JP/KR/IN geographic cluster. The system then generated a MedWatch 3500A form and a PSUR section, fired a Slack #safety‑critical alert, created a Jira P1 ticket, and logged every action; SENTINEL detected a 15× volume spike in that run.

The prototype emphasizes in‑index analytics: the ANALYST agent implements the WHO PRR calculation entirely in ES|QL using STATS for counts, EVAL for ratio math, WHERE for thresholding, BUCKET(report_date, 1 week) for temporal clustering, and geo.country_code for geographic aggregation. Workflows orchestrate agent handoffs and asynchronous pipelines, and deterministic temperatures determine when LLM generation is invoked versus strict numeric computation and rule evaluation.

Sutar framed PHAROS against existing operational pressures: the FDA receives about two million adverse drug event reports annually, and pharmaceutical companies are legally required to detect safety signals within 15 calendar days of a serious report. In practice, manual review across FAERS, EudraVigilance, EHRs, and social media can take weeks to months and consumes 40+ analyst hours per signal, creating opportunity for costly delays or missed signals.

He also highlighted the stakes: Merck’s failure to catch cardiac signals from Vioxx resulted in $4.85 billion in settlements, and single missed signals can lead to fines on the order of $100 million–$1 billion. PHAROS demonstrates how combining ES|QL statistical pipelines with LLM‑assisted document generation and multi‑agent orchestration can drastically cut latency, but Sutar stresses it remains a hackathon prototype that requires dataset validation, regulatory‑compliance checks, reproducibility guarantees, and human‑in‑the‑loop controls before any production deployment.