K12 Language Learning Platform — A Multi-Level Franchise English Learning Platform
Upgrading a K-12 English learning platform to a multi-province franchise model: teacher-supervised AI pronunciation scoring, multi-level Province–Partner–School access control, and unified SSO.
System Scale & Design Thresholds
Designed for a multi-level franchise model, with performance thresholds set from the architecture stage onward.
The Challenge
Upgrading a live platform to a multi-province franchise model, with zero tolerance for downtime or historical data loss.
Zero-downtime data restructuring
Converting the content model from fixed "Weeks" to flexible "Lessons" across tens of thousands of students' historical data.
Multi-level access without data leaks
Strict data boundaries between Province – Partner – School as the system expands to more provinces.
AI pronunciation scoring for children
Children's voices are high-pitched with uneven pacing — AI can't be 100% accurate, so it needs teacher oversight.
SSO & account conflict handling
Unifying login while still safely handling accounts that already existed under the old method.
Leaderboards & reporting without slowing the system
Computing rankings and reports over millions of records without affecting students' everyday learning experience.
Feature Map
Read Aloud AI
- Listen → Record → Submit → Reading comprehension
- 3 scoring criteria: pronunciation, intonation, fluency
- Sample feedback bank
- Teacher review (human-in-the-loop)
Multi-Level Access
- 5 standardized role groups
- Entity binding to Province/Partner/School
- Blocked right at the Controller layer
SSO
- Unified login
- Account conflict handling
- Manual merge tool for admins
- Login audit log
Teaching Content
- Converting the Week structure to Lessons
- Course & material library management
Gamification & Reporting
- Star accumulation, leaderboards
- Role-based dashboards
- Periodic pre-computed reports
A Student's Speaking-Practice Journey
Data Restructuring & Safe Multi-Level Access
A Shadow Tables strategy converts the content structure without downtime, paired with a security interceptor that blocks data leaks right at the request layer.
- Shadow tables + supervised auto-mapping for historical data
- Indexing by Province/Partner/School ID
- Time-based range partitioning by academic year
- Unauthorized access blocked at the Controller layer, before it touches the database
Technical Notes
AI Read Aloud & Unified SSO
An asynchronous scoring pipeline pairs enterprise-grade AI with teacher review, alongside an OIDC SSO flow that safely handles account conflicts.
- Recordings uploaded via presigned URL, processed in the background via a message queue
- A dedicated AI Speech worker, isolated from the main API
- An LLM turns raw scores into natural-language feedback
- Teacher review before scores are published (human-in-the-loop)
- OIDC SSO with a manual account-merge tool for admins
Technical Notes
Technical Architecture
Backend
Frontend
Data & Cache
Infra & Integrations
"Upgrading a live education platform serving tens of thousands of real users demands a completely different mindset than building new: every data-model change has to be designed so nobody notices the wheels were changed while the car was still moving."— The XO Engineering Team