zivAI is an AI teacher and student assistant for O-Level education. The product idea is simple: help teachers manage learning workflows and give students better feedback, even in environments where internet access is not always reliable.

The technical challenge is bigger than a normal dashboard. The system needs learning management features, AI-assisted marking, local school operation, and cloud synchronization.

Core Product Workflows

The platform is built around school workflows:

  • Class and learner management.
  • Assessment creation.
  • Student submissions.
  • Teacher review.
  • Automated short-answer grading.
  • OCR-assisted marking.
  • Reporting and progress tracking.

The goal is not to replace the teacher. The goal is to remove repetitive work and make feedback faster.

Backend Structure

The LMS backend uses Spring Boot and PostgreSQL. This layer owns the core business logic: users, roles, classes, assessments, submissions, review states, reports, and synchronization rules.

For the AI layer, Flask services expose model-backed capabilities such as grading, OCR processing, and knowledge tracing. Keeping those services separate makes the system easier to reason about. The LMS remains responsible for workflow integrity, while AI services focus on inference tasks.

Cloud-Edge Deployment

The platform is designed for hybrid cloud-edge deployment. A school can run an edge node locally and continue operating when connectivity is weak. When internet access returns, the edge node can sync with the cloud.

This matters because education software cannot assume perfect infrastructure. A useful system has to respect the environment where it will actually run.

AI Serving

Models were trained with Huawei ModelArts and MindSpore. For offline serving at the edge, MindSpore Lite makes the AI layer more practical because it reduces dependence on constant cloud access.

That design keeps AI useful without turning it into a network-only feature.

Lesson Learned

The biggest lesson from zivAI is that AI product engineering is not only about the model. The model is one part of the system. The real work is in the workflows, backend boundaries, deployment model, data quality, and user experience around the AI output.