Understand risk
Screening models read structured answers, not guesses.
Self-assessments use form responses, age, vitals, history, and risk factors to return a clear risk band and next-step guidance.
Health AI
Clinical intelligence that supports the patient journey.
Moesha and the AfyaRekod model layer organize risk screening, symptom extraction, vitals analysis, imaging workflows, and governed recommendations with clear safety boundaries.
Model groups5
Serving modes3
GovernanceRequired
AI model breakdownReal-time CPU endpointsAsync GPU for imagingNear real-time vitalsRAG guardrailsModel cardsClinical validation
Understanding the system
Health AI is not one model. It is a set of guided clinical workflows.
Each workflow is separated by what the patient is doing, how quickly the answer is needed, and how much clinical governance is required before an output appears.
Understand language
Text models turn symptoms and records into usable context.
The symptom extractor identifies medical keywords, body areas, duration, severity, and red flags from patient notes and imported records.
Understand change
Vitals models watch trends across diary and device readings.
Pulse, blood pressure, temperature, SpO2, glucose, and diary patterns can be reviewed for trend shifts and escalation thresholds.
Understand images
Imaging and visual analysis run through slower governed paths.
Camera, wound, skin, scan, and radiology workflows can be queued for heavier review where auditability matters more than speed.
Patient experience
What the user should feel: explain, prepare, and escalate.
The AI layer should make the record easier to act on. A patient can ask about symptoms, review a result, prepare for a visit, or understand when a clinician should be involved.
See dashboard flow
Record explanationPlain-language summaries from labs, notes, and history.
Diary and vitals reviewPattern recognition across pulse, BP, SpO2, temperature, and symptoms.
Care escalationRed flags, visit preparation, and clinician-ready summaries.
Capability groups
Each AI capability has a different serving path.
The page separates patient-facing support from heavier clinical workflows so the product feels useful, explainable, and careful.
Real-time CPU endpoint
Self-assessment risk models
Structured risk screening for diabetes, hypertension, stroke, cancer, and TSC depression survey workflows.
Patient-facingDiabetes riskHypertension riskStroke riskCancer riskTSC depression survey
Async GPU or batch
Imaging anomaly and segmentation
CT, MRI, X-ray, and ultrasound analysis should run asynchronously because imaging is heavier and GPU-expensive.
Clinician-assistedX-ray anomalyCT triageMRI segmentationUltrasound review
Near real-time governed endpoint
Predictive and forecasting models
Diagnosis, prognosis, vitals analysis, and recommendations are routed through governed clinical endpoints.
Care pathway supportVitals analyzerDiagnosis supportPrognosisCare recommendations
LLM/RAG and batch cleanup
Medical keyword extractors
Extract symptoms, diseases, and body-part context from patient text and imported clinical records.
Records and symptom flowsSymptom extractionDisease termsBody part detectionRecord cleanup
Guardrailed RAG experience
Moesha patient assistant
Moesha explains records, prepares questions, and guides next actions without replacing a qualified clinician.
Patient guidanceLab explanationVisit prepDiary summarySafety escalation
Real-time vs async
Serve fast only when someone is waiting.
This keeps the patient experience responsive while reserving heavier compute for imaging, segmentation, coding, and batch analytics.
Diabetes, hypertension, stroke, and cancer self-assessment
YesThe user is waiting in the app after submitting a form.
Managed CPU endpoint or lightweight inference serviceSymptom extractor
Yes for app, batch for cleanupNeeded during the symptom flow and later for record normalization.
Guardrailed LLM/RAG service with clinical retrievalDiagnosis, prognosis, and vitals analyzer
Near real-timeClinician and patient workflows need low-latency but governed output.
Governed inference endpoint, stream processor, and online feature storeImaging anomaly detection and segmentation
Usually asyncImages are heavier, GPU-expensive, and rarely require sub-second output.
Async GPU inference or scheduled batch processingTSC depression model
After survey submitSurvey-based risk screening should run after form submission, not continuous monitoring.
CPU endpoint or background workerGovernance
Clinical AI needs proof before production.
Model cards, interpretability, fairness review, and validation gates keep Health AI trustworthy and auditable.
Clinical validation gate
Every model needs signed clinical review before it is promoted to production use.
Interpretability
SHAP, LIME, attention maps, or equivalent explanations must be available for governed outputs.
Drift monitoring
Input distribution, prediction drift, and embedding drift are monitored before retraining is triggered.
Safety boundaries
Patient-facing AI explains and guides; it does not diagnose or replace a qualified clinician.
Moesha guides the patient without replacing the clinician.
Health AI should explain, prepare, summarize, and escalate. It should stay grounded in records, guidelines, consent, and clinical governance.
Open prototype