MIRA: A WhatsApp-Deployed, Rule-Overridden Multi-Agent Pipeline for Pre-Contact Clinical Triage in Low-Resource, Multilingual Settings

MIRA is a five-node pipeline (Intake → Language → Diagnostic → Triage → Referral) that accepts a WhatsApp message describing symptoms, in the patient's own language including romanized Hindi and Telugu, and returns a triage decision (Emergency / Urgent / Routine) plus a referral. It is explicitly framed as a triage decision-support tool, not a diagnostic agent: a deterministic red-flag rule layer can override the underlying language model unconditionally, and every node's input/output is logged verbatim for audit. This is an interim evaluation on a 30-case synthetic benchmark. Two independent practicing physicians, both co-authors, reviewed the case bank: the first (the system's clinical co-advisor) reviewed system output alongside each case; the second reviewed all 30 cases fully blind, with no system output and no visibility into the first reviewer's assessments. Of 29 cases scored against the first rater, MIRA matched the clinically confirmed triage tier in 14 (48.3%); of 27 cases scored against the second, independent rater, MIRA matched in 10 (37.0%). The two physicians agreed with each other on 22 of 28 directly comparable cases (78.6%), and every disagreement ran in the same direction, with the second rater systematically more cautious than the first. This report also documents findings considered more important than any single headline number: a missed stroke presentation where the deterministic rule layer did not fire, one confirmed dangerous under-triage of a pediatric respiratory danger sign, and a no-decision rate (37.9% against the first rater's scored set, 48.1% against the second) where the system produced no triage output at all rather than a conservative default — including on all three of its own romanized-language test cases. This is a system evaluated on synthetic cases by two independent clinical raters, not a validated clinical tool. Full methodology, limitations, and findings are in the attached report.