Technology & infrastructure

AI infrastructure
built for clinical trust.

Four AI models, AWS healthcare-grade infrastructure, and NVIDIA GPU acceleration — working together to deliver triage decisions clinicians can rely on and patients can trust.

AI technology

AI isn't a feature here. It's the entire care pathway.

Four models work in concert — each audited, each with a human reviewer in the loop, each improving weekly on confirmed outcome data.

01

Multilingual symptom understanding

A fine-tuned large language model trained on de-identified consultation transcripts from African clinical settings and WHO clinical guidelines. The model conducts structured symptom intake in English, Pidgin, Yoruba, Hausa, Igbo, and Swahili — adapting follow-up questions based on prior answers.

Amazon BedrockFine-tuned LLMNLPMultilingual
02

Risk stratification engine

A gradient-boosted ensemble model that scores symptom severity across 14 clinical axes — including 23 red-flag patterns associated with cardiac events, stroke, sepsis, and obstetric emergencies. Output is a four-tier risk classification reviewed by a CMO-approved decision framework.

Gradient boostingClinical decision support14-axis scoring
03

Clinician matching via vector search

Practitioner embeddings encode specialty, sub-specialty, languages spoken, real-time availability, and aggregated patient outcome scores. FAISS-powered vector similarity search returns a ranked shortlist in under 200ms — updated daily as clinician availability changes.

Vector searchFAISSEmbeddingsReal-time matching
04

Continuous outcome learning

Every confirmed diagnosis and clinician note is fed back into the training pipeline — de-identified and consented. Models are retrained weekly. Each retraining cycle is audited against a held-out clinical benchmark before deployment.

Feedback loopWeekly retrainingAudit trailNDPR compliant

Cloud — AWS

Healthcare-grade cloud backbone.

Every AWS service in our stack was chosen specifically for healthcare compliance — HIPAA-eligible, NDPR-compatible, with audit trails on every access event.

Amazon Bedrock

Managed LLM inference with no model weights leaving the VPC. Powers symptom understanding and multilingual response generation.

AWS Fargate (ECS)

Containerized backend APIs — autoscaling, no server management, isolated per-service security boundaries.

Amazon RDS (PostgreSQL)

Encrypted relational store for patient records. Field-level KMS encryption. PITR backups with 35-day retention.

AWS Lambda

Async clinical workflows: risk-score post-processing, notification dispatch, model feedback ingestion.

Amazon S3 + KMS

De-identified training data storage with server-side encryption and object-level access logging.

Amazon CloudFront + WAF

Global CDN edge delivery with managed WAF rules — rate limiting, bot detection, SQL injection and XSS protection.

Amazon Cognito

MFA-enforced authentication with PKCE OAuth flows. No passwords stored in application code.

AWS CloudWatch + GuardDuty

Full audit logging on every API call. GuardDuty detects anomalous access patterns and potential data exfiltration.

AI acceleration — NVIDIA

Real-time inference. No queue.

Triage decisions need to feel instant. NVIDIA's inference stack lets us serve multiple concurrent AI models with sub-100ms latency — at the scale Africa demands.

NVIDIA Triton Inference Server

Multi-model GPU serving for concurrent inference requests. Supports TensorRT, ONNX, and PyTorch backends on a single server fleet.

TensorRT

Quantized and optimized model graphs for sub-100ms triage inference at peak load — even on single-GPU instances.

NVIDIA NeMo

Multilingual speech recognition for Pulsa's voice triage channel. Fine-tuned on West African English and Pidgin phoneme datasets.

NVIDIA MONAI

Medical imaging AI framework on the roadmap for radiology report summarization and imaging-informed triage.

NVIDIA NIM Microservices

Portable, secure model deployment containers that keep inference workloads cloud-agnostic and reproducible across regions.

Security & compliance

Data protection is the product.

AES-256 at rest

Field-level KMS encryption on all patient records. Keys rotated annually.

TLS 1.3 in transit

All API calls encrypted end-to-end. No cleartext health data on the wire.

NDPR aligned

Nigeria Data Protection Regulation compliant. Data subjects can export or delete at any time.

HIPAA-equivalent

Controls mapped to HIPAA Security Rule. Annual third-party audit planned for Series A.

Want a technical deep-dive?

We run 60-minute technical sessions for engineering and security teams. Request one and we'll send a calendar link within 24 hours.

Request a demo →