Chinese Herbal Recognition Platform
Chunxiao Technology · 2019
Role: Platform Designer & AI Workflow Developer
Cover
End-to-end AI platform for image classification: category management, data annotation, model training (YOLOv4/ResNet/MobileNetV3), evaluation with release-gate criteria, and inference serving. Designed as reusable workflow for multiple visual classification domains.
Full-loop annotation-to-inference pipeline, reusable for multiple classification domains
Problem
Traditional Chinese medicine identification needs a systematic AI platform covering the full workflow from data annotation to model serving.
Solution
Closed-loop AI platform: category creation → image annotation → quality review → model training with experiment tracking → evaluation with release gates → versioned model publishing → inference API.
Key Highlights
- ▸Designed end-to-end AI platform workflow from category setup to recognition inference
- ▸Built image annotation process with quality review and dispute handling
- ▸Added model training pipeline with experiment tracking and model registry
- ▸Low-confidence sample feedback to re-annotation queue for continuous improvement
Tech Stack
What I Learned
AI platforms need standardized release criteria for model publishing; low-confidence feedback loops improve model quality over time; designing for extensibility enables quick domain adaptation.