AI LMS

Reconstructing a legacy system to create an AI-powered Teaching SaaS for international schools. Key achievements include a zero-downtime migration and supporting four R&D lines with a single design resource.

Exam Assembly Workflow Reconstruction

Problem: A rigid 4-step pipeline (Select → Score → Comment → Report) with complex validation rules and frequent errors. Scoring was critical but so infrequent that every use meant relearning.
Solution: Three integrated strategies — a "Shopping Cart" for one-click question selection, Smart Assembly with drag-and-drop and auto-scoring, and Optimized Scoring with built-in templates reducing 90% of manual setup.
Result: Assembly time dropped from 15 min to 5 min (60%+ gain). This became the first module to be fully migrated and stabilized.

AI Grading — Landing AI Where It Matters Most

Why here: Language training is dominated by subjective questions (speaking, listening, writing) — the highest-leverage workflow for AI optimization.
Design strategy: A "Negative Operation" logic — assume AI is correct by default, teachers only reject errors rather than confirming every answer. Combined with an Immersive Sidebar for high-frequency student/task switching, and AI Traceability so teachers can inspect the original evidence behind every AI judgment.
Result: Grading time from 20 min to 5 min (80% gain), 90% AI adoption rate. A co-creation partner committed to migrating all annual teaching tasks to the platform based on this feature.

Design Ops — Driving Efficiency Through Design

1 vs 4 Leverage: A single Design System serving 4 dev teams — unified module-level delivery with upfront design reviews.
1-to-N Abstraction: Abstracted the core "Assign → Take → Grade" architecture so expanding from TOEFL to IELTS took just 1 week of prep, with >70% configuration reuse.
Agile Delivery Loop: Built a WeChat + DingTalk direct feedback channel between clients and the product-design team, achieving D+1 delivery for urgent requests.