MediSure(Authentica)

01 · The problem

Counterfeit and tampered pharmaceutical products are a critical public health issue in African markets, particularly Nigeria. Manufacturers, regulators (NAFDAC), and supply-chain participants needed a way to authenticate products at the point of inspection using just a smartphone camera — no specialized hardware.

02 · What I built

The complete user-facing Flutter app: real-time camera scanning with a pixel-change detection system that auto-captures four frames as you rotate the product, eight TFLite models running on-device, an offline-first SQLite architecture with background sync, and a 3-in-1 role-based admin app serving Admin, Manufacturer, and NAFDAC users. I also trained the YOLOv8/v11 and EfficientNet models on Colab, expanding datasets 4× with synthetic augmentation.

03 · The impact

88.5%–95.3% model accuracy across four product categories, 99.9% app uptime via adaptive local/cloud inference, 25% fewer false positives from the pre-processing pipeline, and 60% less manual entry thanks to ML Kit OCR. Live admin panel at admin.authentica.africa.

What I did

• Engineered an adaptive AI inference system maintaining 99.9% uptime by toggling between local TFLite models and cloud APIs based on real-time network latency.
• Built a self-optimizing image pre-processing pipeline that reduced false-positive predictions by 25% through motion-blur analysis before inference.
• Trained 8 models (YOLOv8, YOLOv11, EfficientNet V2-B2) on Colab T4 GPUs to 88.5%–95.3% accuracy using ~8,000 images, expanding datasets 4× with synthetic augmentation.
• Built a pixel-change detection system that auto-captures four inspection frames as the user rotates the product — hands-free multi-angle scanning.
• Integrated ML Kit OCR for batch numbers and expiry dates, cutting manual input time by 60%.
• Architected a 3-in-1 role-based admin app (Admin, Manufacturer, NAFDAC regulator) with API-driven role detection and page-level access control.
• Designed offline-first SQLite architecture with scheduled background sync to the Python/MySQL backend.