Materna — AI-Powered Obstetric Safety Review

Materna is a multi-agent obstetric safety review system built for the Band of Agents Hackathon . Four specialist AI agents — Intake, Dating & Risk, Guideline, and Auditor — coordinate through Band's Agent API and shared rooms to review antenatal cases. Dr. Saima Javed, a practicing gynecologist, serves as the human-in-the-loop gate through Band's Human API, holding final authority on every escalated case. HOW IT WORKS: 1. Intake Agent normalises handwritten or typed clinical notes (English and Urdu supported) into structured JSON using AI/ML API and Gemini Vision. 2. Dating & Risk Agent computes gestational age via the Hadlock formula, detects LMP-USG discordance, and fires deterministic risk flags for pre-eclampsia, gestational diabetes, anaemia, and dating errors. 3. Guideline Agent checks the structured case against the antenatal care rulebook and issues a compliance veto when required investigations are missing. 4. Auditor Agent adversarially reviews the Guideline Agent's output — challenging borderline decisions and catching missed flags before anything reaches the human reviewer. 5. The escalation decision (must_escalate) is a 13-line pure function — no LLM can influence whether a case reaches Dr. Saima. This has been proven through 132 adversarial hardening tests. 6. Dr. Saima reviews flagged cases, approves or overrides the AI treatment plan, and only then is the review packet cryptographically sealed with a SHA-256 hash chain. 7. A tamper-evident clinical PDF is generated via ReportLab with the final hash embedded in the footer. CLINICAL IMPACT: In Pakistani antenatal clinics, overworked staff and handwritten records mean critical risks go undetected. Materna catches what manual review misses: pre-eclampsia detection rises from 34% to 98%, GDM screening compliance from 28% to 99%, and anaemia flagging from 41% to 97%. Each case review saves 39 minutes of clinician time — equivalent to 6 additional patients seen per clinic per day.