Bongekilenkosi/BIZUSIZO

BIZUSIZO — hybrid deterministic-AI triage system for South African primary healthcare. Code, validation data, and paper artefacts supporting the medRxiv preprint. # BIZUSIZO **Hybrid deterministic–AI triage system for South African primary healthcare, delivered via WhatsApp and aligned to the South African Triage Scale (SATS).** This repository contains the code, validation data, and scientific artefacts supporting the preliminary safety validation of BIZUSIZO. It accompanies the preprint under revision at medRxiv (MS ID 349781) and is intended for reviewers, replicators, and researchers working on clinical AI for low- and middle-income country primary healthcare. --- ## What BIZUSIZO is A WhatsApp-delivered pre-arrival triage tool that combines: - **AI-assisted classification** (Anthropic Claude) for free-text symptom descriptions in any of South Africa's 11 official languages. - **A Deterministic Clinical Safety Layer (DCSL)** — a rule-based engine that overrides AI output for 53 clinical discriminator categories (14 RED life-threatening, 19 ORANGE very-urgent, 20 YELLOW urgent). The DCSL runs independent of AI availability and catches coded emergency presentations even when the AI service is down. - **Continuous integration as clinical safety infrastructure** — a 121-vignette multilingual regression test runs on every deployment and blocks release on any safety failure. The design premise is that AI safety in resource-constrained primary healthcare is best treated as a systems-engineering problem (layered deterministic constraints, multi-agent verification, continuous monitoring) rather than a model-accuracy problem. --- ## Security posture The DCSL rule engine, rule names, and architecture are fully published in this repository. The **operational keyword sets** (the specific phrases in all 11 languages that cause each rule to fire) are maintained privately and loaded at runtime from `config/dcsl_keywords.json`, which is not committed here. `config/dcsl_keywords.example.json` shows the required JSON shape with empty arrays. This follows established practice for safety-critical clinical AI systems: publish the ar …