Overview

Mathru explores how a familiar WhatsApp conversation could help expectant mothers stay connected with their assigned Public Health Midwife (PHM). Built for the IDEALIZE 2026 mini-competition, it is a software prototype, with clinical review and validated health data still outstanding.

The Python application uses AgentKernel's WhatsApp adapter and OpenAI Agents integration. Five agents separate role resolution, registration, clinic-schedule queries, symptom-report handling, and the midwife workspace.

A shared channel, distinct workflows

Mothers can register their MOH division, expected delivery date or child's birth date, and assigned midwife. The PHM workflow supports caseload queries and escalation acknowledgement. Phone-linked identity and an operator-approved midwife registry determine which workflow and records are available.

Python tools handle structured records and date calculations; agent handoffs route the conversation between focused responsibilities. SQLite persists mother and escalation records, while conversational sessions use separate in-memory storage.

Escalation with context

The escalation path attempts to notify an approved assigned midwife with the mother's report and context. Failed or unapproved deliveries are recorded as undelivered, with a direct-care fallback message. WhatsApp webhook verification and server-resolved identity sit at the integration boundary.

Prototype boundaries

The bundled clinical tables are placeholders. Schedule tools withhold unverified dates, and the current screening tool treats every nonempty symptom report it receives as a red escalation. That behavior demonstrates routing and delivery; it is not validated symptom assessment, and model routing does not guarantee that every report reaches the tool.

Diagnosis and medication restrictions, moderation, and response checks are development controls, not evidence of clinical safety. The next steps are verified source data, clinical review, and stronger end-to-end reliability before considering real use.

What the prototype explores

The interesting engineering boundary is between a conversational interface and accountable operational state: resolving identity, handing off between agents, recording an escalation, and making delivery failure visible rather than implying that a message reached a clinician.