How AI Drives Education: Inside Zuoyebang’s Massive Question Bank & Smart Grading

In this interview, Wang Yan, chief architect of Zuoyebang, discusses the challenges and opportunities of applying artificial intelligence in education, emphasizing the need to bridge the gap between hype and real‑world implementation.

PART 01 – Education Product Foundation: The Three‑Step Question Bank

Zuoyebang’s question bank, now over 540 million items, was built on three key pillars:

1. Leverage early advantages : Originating from Baidu’s internal Q&A community, Zuoyebang assembled full‑time researchers and part‑time teachers to create the largest online question‑production platform.

2. Value‑driven crowdsourcing : By breaking the bank into small, independent tasks, the crowdsourced workflow ensures rapid, comprehensive growth while maintaining quality.

3. Rich metadata and knowledge graph integration : Each question is tagged with knowledge points, difficulty, and dependencies, linking to a knowledge graph that enables efficient retrieval and recommendation.

AI has progressively automated many steps, from image capture and OCR to auto‑tagging, formula formatting, and error correction, cutting manual effort and boosting accuracy, which helped reduce response time from eight seconds to one second.

PART 02 – Automated Grading: How Machines Solve Graphic Problems

For subjective questions, especially graphic ones, Zuoyebang combines OCR, NLP, and knowledge‑graph analysis to recognize student input, detect logical errors, and generate personalized feedback reports.

The system runs on a cloud‑native, multi‑cloud disaster‑recovery architecture, guaranteeing high stability even under massive concurrent usage.

Key features include:

Step‑level automatic grading that reduces teachers’ workload and improves student outcomes.

Vectorized feature extraction for graphic questions, enabling accurate matching of shapes and diagrams.

OCR‑based exam‑paper restoration that corrects distorted scans in seconds, also useful for digital submission polishing.

Knowledge‑graph‑driven recommendation that links questions to related concepts for deeper learning.

PART 03 – From Digitization to AI: Respecting Existing Habits

Digitizing textbooks, blackboards, and PPTs is essential for AI to operate on educational content. However, Zuoyebang emphasizes preserving teachers’ and students’ familiar workflows, such as retaining paper‑based exams through the “original sheet trace” feature.

AI also supports new use cases like real‑time heart‑rate monitoring in physical education, automatic jump‑rope counting, and posture correction via computer‑vision.

PART 04 – Perspectives on the AI Industry

Basic research vs. frontier tech : Fundamental research provides a scalable base, while frontier innovations (e.g., large models, multimodal AI) can unlock new capabilities; both deserve investment.

Model size and accessibility : Massive models improve accuracy marginally but require prohibitive compute resources, limiting widespread adoption. Cost‑effective, high‑performance solutions are more valuable for mass education.

Finding AI landing opportunities : Identify existing technical assets, match them to concrete educational scenarios, run pilots, and iterate. B‑side (school) customers demand higher precision than C‑side (individual) users.

Overall, Zuoyebang’s strategy combines deep AI research with pragmatic engineering to deliver scalable, affordable AI services that enhance learning without disrupting established habits.