Sharing technical practices from Zuoyebang
This article presents a method to improve keyword detection in large‑scale speech recognition by integrating a prefix automaton into the beam‑search decoding of seq2seq models, enabling real‑time addition of new terms while reducing computational overhead compared to traditional approaches.
This article explores the motivations, challenges, and design principles behind adopting a multi‑cloud active‑active architecture, emphasizing how it enhances stability, reduces costs, and improves efficiency, while detailing practical solutions for networking, compute, containers, service discovery, traffic routing, and data storage in a cloud‑native environment.
The first Cloud‑Native Practitioners Conference brought together engineers from Zuoyebang, Zhihu, Zhuanzhuan and others to share large‑scale Kubernetes, serverless and Istio experiences, industry trends, and practical solutions that drive stability, efficiency and digital transformation.
Zuoyebang’s chief architect Wang Yan explains how the company leverages AI, large‑scale question banks, automated grading, knowledge graphs, and cloud‑native infrastructure to transform education, detailing the three‑step question‑bank construction, AI‑assisted marking of graphic problems, and strategies for practical AI adoption in schools.
This article describes the design and implementation of an automated FlinkSQL tuning system that monitors metrics, evaluates task health with rule‑based logic, calculates optimal resource adjustments, and performs fast scaling to reduce cluster waste, lower operational costs, and maintain SLA compliance.
The 2022 Cloud Native Industry Conference in Beijing, organized by the China Academy of Information and Communications Technology and the China Communications Standards Association, showcased 14 exemplary cloud‑native cases—including Zuoyebang’s search system—highlighting the rapid growth of China’s cloud‑native ecosystem, its technical innovations, and the release of a national cloud‑native security testing platform.
This article examines the challenges of English pronunciation teaching in China and presents a practical end-to-end phoneme‑level mispronunciation detection system that leverages CTC models, attention‑based text fusion, and data augmentation to dramatically reduce false alarms and improve diagnostic accuracy.
Facing fragmented devices, high memory usage, and unreliable signaling, the Zuoyebang live‑streaming frontend was redesigned with container‑based isolation, a modular SDK, and state‑synchronization mechanisms, resulting in timely memory reclamation, reduced CPU load, and dramatically lower black‑screen incidents.
Zuoyebang’s data team replaced fragmented, slow query solutions with Apache Doris, building a unified real‑time data warehouse that dramatically cut query latency from hours to seconds, streamlined data modeling, and improved reliability across diverse business scenarios, while integrating with Flink, Kafka, and ES via a unified API.
This article describes the challenges encountered when migrating tens of thousands of Kubernetes cronjobs from VMs to a cluster—node instability, low resource utilization, and scheduling delays—and explains how introducing a serverless architecture with virtual nodes, a custom job scheduler, unified logging and monitoring, and sandbox reuse restored stability, improved performance, and reduced resource costs by about 70%.
