Phase 2: Improve Model Accuracy and Deploy to Cloudflare Pages

[deepdevjose]

Objective

Improve digit recognition quality beyond the MVP baseline and automate deployment to Cloudflare Pages using GitHub Actions.

Current Context

• The MVP already supports drawing, preprocessing, inference, and confidence display.
• A stronger trained model, better evaluation, and deployment automation are still needed.

Scope

• Improve model quality (architecture, training, and TensorFlow.js export).
• Improve preprocessing robustness for real canvas input.
• Improve prediction UX (top-3 classes, confidence states, clearer errors).
• Add CI/CD workflow for Cloudflare Pages deployment.

Out of Scope

• Dedicated backend services.
• Multi-language UI support.
• Advanced end-user analytics.

Acceptance Criteria

• Validation and test accuracy improve against a documented baseline.
• Production uses a trained TensorFlow.js model artifact.
• Deployment to Cloudflare Pages runs automatically on push to main.
• CI validates model assets before deployment.
• Manual checks for digits 0, 1, 7, 8, and 9 show coherent predictions.

Deliverables

• Trained TensorFlow.js model artifacts for production.
• GitHub Actions workflow for build and deployment.
• Model asset validation script.
• Documentation for training decisions, metrics, and deployment setup.

Dependencies

GitHub Secrets

• CLOUDFLARE_API_TOKEN
• CLOUDFLARE_ACCOUNT_ID

GitHub Variables

• CLOUDFLARE_PROJECT_NAME

Risks

• Training preprocessing may not match browser preprocessing.
• Production model asset paths may fail after deployment.
• Cloudflare token scopes or account settings may be misconfigured.

Definition of Done

• CI build and deploy are green.
• Cloudflare Pages site is updated from main.
• Production can load model.json and all referenced shard files.
• Prediction quality is measurably better than baseline in metrics and manual checks.