React Native + Expo
Profiles, camera flow, verification, trends, consent, and referral UI.
→Technology
Built to read the paper chart without overloading the phone.
NeoScan separates fast on-device document work from server-side computer vision, WHO analytics, secure records, and audit logging.
Architecture
From mobile capture to structured record.
Each layer has a focused job. The diagram presents the specified design, not a claim that production infrastructure is already deployed.
Google ML Kit
Document edges, auto-crop, perspective correction, and squared JPEG output.
→OpenCV
Orientation, contrast, shadow handling, and HSV pen-mark separation.
→YOLOv11
Chart axes, curves, plot points, and handwriting-noise detection.
FastAPI
Authentication, inference orchestration, analytics, flags, and reports.
→WHO LMS
Sex-specific, day-precise reference lookup and Z-score calculation.
→PostgreSQL
WHO tables, patient records, scans, flags, referrals, and audit trails.
→Digital growth view
Verified trends, review flags, offline sync state, and consented sharing.
Why this split
Fast document correction near the camera. Consistent inference in the cloud.
The SRS keeps Google ML Kit work on-device and moves YOLOv11 to FastAPI to reduce heat, memory pressure, and hardware-driven accuracy variation.
- Under 1.5 seconds
Target for local edge detection and rectification after capture.
- Under 3 seconds
Target for upload, inference, WHO calculation, flag evaluation, and response over standard 4G.
- Under 250 MB RAM
Specified ceiling for normal mobile-client operation.
These are design targets from the June 2026 draft SRS, not measured results.
Technical stack
Chosen for the chart, the workflow, and the team.
MobileReact Native + Expo
One iOS and Android codebase, with Expo-managed integration and SVG chart rendering.
Document scanGoogle ML Kit
On-device corner detection, auto-crop, perspective correction, and document-squared image output.
Image processingOpenCV 4.x
Adaptive thresholding, orientation normalization, and HSV masks that separate pen marks from printed curves.
DetectionYOLOv11 / PyTorch
Custom model training for the Kenya MOH chart format and multi-class element detection.
BackendFastAPI + Uvicorn
Python API gateway, authentication, inference, WHO analytics, anomaly rules, logging, and report generation.
DataPostgreSQL 15+
Relational storage for reference tables, growth records, consent, referrals, immunization requirements, and immutable audit events.
Computer vision pipeline
Designed around the printed geometry of one chart family.
The model is constrained to the Kenya MOH Road-to-Health chart format. A different print layout requires coordinate recalibration and potentially new model training.
- Normalize
Read EXIF and structural anchors, then rotate upright.
- Flatten
Correct skew, crop the chart, and mitigate lighting artifacts.
- Separate
Mask blue or black pen marks before object detection.
- Detect
Find axes, percentile curves, plot points, and text noise.
- Map
Translate local pixel positions into age and measurement values.
Tightly coupled: The SRS states that OpenCV color separation and YOLO detection must both run because the model cannot reliably distinguish handwriting from printed curves on its own.
Offline queue
Encrypted SQLite holds scans and metadata temporarily.
Capture→AES-256 local store→Network monitor→FastAPI sync
Synchronization architecture
Chronology survives connectivity gaps.
When several scans are queued, the backend is specified to process them in time order and preserve the original capture time for flags and audit history.
Images are resized to a maximum width of 1920 pixels and targeted at 80–85% JPEG quality to protect grid-line detail on limited bandwidth.
Security architecture
Protection across device, network, API, and database.
Layered controls
The SRS defines these as architecture requirements under Kenya’s data protection framework.
Local storageAES-256 encrypted SQLite; Keychain or Keystore for tokens.
TransportHTTPS over TLS 1.3 on port 443.
API accessShort-lived JWTs with rotating refresh tokens.
DatabaseAES-256 at rest, RBAC boundaries, and immutable logs.
Evidence behind the architecture