CareMate is built from the ground up to be intelligent. A clinical knowledge graph, three specialised LLM-powered agents, and an AI-native EHR that work together so the system reasons, remembers, and guides — not just retrieves. Every layer has AI at its core. Nothing is bolted on.
Most clinical decision support tools are search engines with a medical skin. They wait for you to ask the right question, then return a list of results. CareMate is different. It reasons over a structured knowledge graph to generate differential diagnoses. It guides treatment step by step. It remembers every patient encounter and uses that history to make the next one better. Intelligence isn't a feature — it's the architecture.
Traditional tools:
Nurse types symptom → gets a list of articles
Nurse types drug name → gets a dosing table
Patient returns → starts from scratch
CareMate:
Nurse describes complaint → AI returns ranked differential + triage acuity + safety review
Diagnosis confirmed → AI walks through treatment protocol + prescribing + discharge
Patient returns → AI has already read full history, medications, care gaps
Across low- and middle-income countries, frontline primary care workers face interconnected challenges that no single point solution can address.
Hundreds of pages of clinical protocols indexed by diagnosis — which requires knowing the answer before asking the question. No system reasons from symptoms to conditions.
Patients see a different nurse every visit. Paper records get lost. No system remembers the patient and alerts the clinician to what matters.
From diagnosis to prescribing to discharge: ad hoc. Drug interactions unchecked, care gaps invisible. No system guides the encounter step by step.
Weekly chart audits after the damage is done. No system provides real-time visibility into clinical decision-making.
Rural clinics bear the highest disease burden with the fewest diagnostic tools. No system answers clinical questions on demand from trusted sources.
Triage, treatment, records, and oversight exist as separate tools that don't share data. No system integrates the full arc of a patient encounter.
Not five separate tools stitched together. One platform where a knowledge graph, three LLM-powered agents, and an AI-native EHR share context and reason together.
The structured intelligence layer. National treatment guidelines transformed into a typed knowledge graph — not flat text, but a web of relationships between conditions, symptoms, medicines, danger signs, and referral criteria. Semantic embeddings enable fuzzy matching. Patient-language synonyms bridge the gap between how patients talk and how guidelines are written. Every agent reasons over this graph.
AI that reasons from symptoms to conditions. Extracts clinical features from free-text complaints using an LLM, expands them through the knowledge graph, scores conditions across multiple dimensions (symptom match, prevalence, demographics, vitals), and returns a ranked differential with nationally standardised SATS triage acuity — all safety-reviewed.
LLM extraction + graph reasoning + deterministic scoring
AI that guides treatment decisions. Walks the nurse from confirmed diagnosis through the guideline treatment protocol: first-line medicines with weight-based dosing, non-pharma interventions, referral criteria, and discharge planning. Checks drug interactions against patient history. Flags stock-outs and suggests STG-compliant alternatives.
Structured workflow + medication intelligence
AI that answers clinical questions instantly. Retrieval-augmented generation (RAG) grounded in guideline source text. Medication dosing, red flag checks, patient education — cited back to the exact STG paragraph. Like having a clinical reference librarian who has read every page of every guideline, on call 24/7.
RAG + semantic search + source citation
Not just a record system — a context engine for the agents. A lightweight EHR that gives every agent longitudinal patient context. The Triage Agent doesn't ask "what's wrong?" — it asks "what's wrong given that this patient has diabetes, was seen 3 months ago for the same complaint, and had a rash from Amoxicillin?" AI-generated visit summaries tell the next clinician what happened and why it matters. Care gaps are surfaced proactively. The system gets smarter with every encounter.
From the moment a patient walks in to their next follow-up visit, every step is AI-assisted.
Nurse enters complaint and vitals. The Triage Agent reasons through the knowledge graph, returns a ranked differential with SATS acuity colour, danger signs, and targeted assessment questions — in under 10 seconds.
The Encounter Agent walks through the guideline protocol: first-line medicines with dosing, non-pharma interventions, referral criteria. Drug interactions checked against the patient's medication history.
The Clinical Assistant answers instantly from guideline source text via RAG. "What's the paediatric dose for Amoxicillin?" — cited, sourced, and context-aware.
The EHR generates an intelligent visit summary and persists the encounter. Next visit — even with a different nurse — CareMate shows the full history, surfaces care gaps, and alerts to continuity risks.
Every architectural decision optimises for clinical accuracy, speed, and safety in low-resource environments.
Purpose-built AI agents using Anthropic's Claude models with structured tool use. Not prompt-and-pray — deterministic pipelines with LLM reasoning at targeted steps: symptom extraction, clinical re-ranking, and safety review.
National treatment guidelines transformed into a typed graph with 12,050 edges. Batch CTE queries, synonym expansion, and multi-dimensional scoring — no vector search needed. Graph reasoning outperforms embedding-only approaches.
Clinical Q&A grounded in 1,532 knowledge chunks with 512-dimensional semantic embeddings. Every answer cites the exact guideline paragraph. Hallucination-resistant by design — if it's not in the guidelines, it's not in the output.
Defence-in-depth: every triage output passes through an independent LLM safety review. Danger signs trigger automatic escalation. Medication clashes blocked. Acuity can only go up, never down, through the pipeline.
Under 10 seconds end-to-end in production. Batch CTE queries replaced per-term lookups (32s → 2s). Deterministic synthesis eliminated unnecessary LLM calls. Parallel async execution across all independent operations.
Country configuration model with source-tagged knowledge bases, pluggable triage systems (SATS, ETAT, custom), adaptive patient IDs, and per-country formularies. Add a country's guidelines — get a CareMate deployment.
Validated with real clinical vignettes across multiple domains by an independent clinician with 20 years of SA public health experience.
Every target condition found in top-5 results across 92 clinical test cases spanning the full breadth of the STG knowledge base.
30 vignettes by Dr Tasleem Ras across 6 domains (Pregnancy, Under 5, Schoolgoing, Adolescent, Adult, Geriatrics). 24/25 correct Top-1 when the knowledge base covers the condition.
South African Triage Scale with TEWS vital sign scoring (7 components), clinical discriminators, and age-stratified thresholds. Not a homegrown system — the national standard, computed deterministically.
Every missed case in Phase I validation was traced to a knowledge base gap (condition not in STG), not an algorithm failure. When the knowledge exists, the AI finds it.
Every recommendation traces to a national treatment guideline. The AI reasons over structured clinical knowledge, not generic medical text from the internet.
Designed for 5-minute consultations in nurse-led clinics. Complaint in, differential out, protocol applied. Not a doctor's EHR shoehorned onto frontline workers.
Patient-language synonyms in isiZulu, isiXhosa, Afrikaans. The AI understands how patients describe symptoms, not just clinical terminology.
Triage, treatment, Q&A, and patient records in one integrated platform where all components share context. No more stitching together disconnected tools.
Country-configurable architecture: clinical guidelines, triage systems, formularies, ID systems, and languages. Built in South Africa, designed for global LMIC deployment.
Independent safety review on every output. Danger sign escalation. Medication clash blocking. Acuity can only go up through the pipeline. Defence-in-depth, not afterthought.
We're looking for clinical partners, research institutions, and health system innovators who believe AI should be built into the foundation of primary care — not bolted on as an afterthought.
Get in Touch →