michflox/mamasafe-ai

Offline-first clinical decision support for maternal emergencies in Africa # MamaSafe AI **Offline clinical emergency copilot for Africa's frontline health workers.** MamaSafe AI is an offline-first, multilingual, AI-assisted clinical workflow and emergency-escalation copilot for frontline healthcare workers in Nigeria — midwives, nurses, CHEWs, physicians, and anesthesia providers working under severe time, staffing, referral, and connectivity constraints. It turns verified clinical guidelines into actionable, step-by-step emergency workflows that help existing health workers recognize emergencies earlier, execute critical protocols more consistently, and communicate referrals more effectively. It is built on the safety-first **AnesthesiaOS Africa** platform lineage and is being developed for a **2026 Google Africa Applied AI Lab** application. The MVP proves one workflow end-to-end: **postpartum hemorrhage (PPH)** — from intake through risk categorization, prioritized checklist, dose/fluid calculation, escalation, structured referral and SBAR handoff, to a tamper-evident audit trail — demonstrated entirely in simulation mode. --- ## ⚠️ Safety and Data Disclaimers > **Development / simulation only.** MamaSafe AI is **not** FDA cleared, **not** CE marked, and **not** approved or registered for clinical use by any authority. It is **not** connected to real clinical devices. It produces clinical decision **support — recommendations, never commands** — and every consequential action requires human clinician confirmation. See `docs/SAFETY.md`. > > **Synthetic data only.** All cases, scenarios, and test data in this repository are synthetic. No real patient data is used, stored, or permitted in this development system. ## Architecture in one paragraph Anything clinically consequential (doses, thresholds, contraindications, scoring, escalation) is computed by a **deterministic, version-controlled, guideline-cited rule engine** — never by the LLM. The **generative layer** (Gemini cloud / Gemma edge, with a mandatory templated offline fall …