Alibaba CloudEducationCustomer Service AutomationAI and Machine LearningGenerative AI / LLMCost ReductionCustomer ExperienceProductivity ImprovementScalability2025Alibaba CloudAC#ff6a00EducationChinaJingzhunxue needed a low-latency, high-stability infrastructure for an AI tutoring experience that mixes real-time audio and video dialog, ASR, TTS, formula recognition, long-context reasoning, and strong safety…automatic expansion can complete within 30 secondsOct 27, 20256Customer Service AutomationApplication Real-Time Monitoring Service (ARMS) · Bailian · Elastic Compute Service (ECS)02025-10-276
Jingzhunxue needed a low-latency, high-stability infrastructure for an AI tutoring experience that mixes real-time audio and video dialog, ASR, TTS, formula recognition, long-context reasoning, and strong safety controls. User traffic spikes heavily at night and on weekends, so the company also needed GPU elasticity and a way to choose and tune foundation models for educational accuracy and content safety.
↗automatic expansion can complete within 30 seconds
↗efficiency improved 40%
↗infrastructure SLA is 99.975%
↗labor cost reduced 35%
Reported impact
Business outcome
Jingzhunxue turned its tutoring product into a real-time AI teaching service that can scale nationally. The platform holds response time to 1.2 to 1.8 seconds, scales within 30 seconds under demand spikes, reduces labor cost by 35%, lifts overall efficiency by 40%, and supports a tutoring experience designed to give many students personalized one-to-one guidance at a much lower delivery cost.
Cost Reduction
Customer Experience
Productivity Improvement
Scalability
Solution
Jingzhunxue built its AI tutor stack on Alibaba Cloud around ECS, GPU cloud servers, SLS, ARMS, QwenMax, and Bailian. The infrastructure layer provides a stable elastic compute base with 99.975% SLA and 30-second scale-out for evening or weekend demand spikes, while SLS and ARMS give full-link observability so the team can identify bottlenecks and trigger what it calls compute self-healing before users feel service degradation. On the model side, the company is an early deep user of QwenMax and uses Bailian to train an education-specific model behind the Hanzue Teacher super-agent. That model stack handles personalized tutoring, AI homework correction, automatic question-bank extraction, and long-context educational dialog while integrating speech and teaching workflows. The result is a production tutoring system that mixes domain-tuned AI teaching capability with elastic infrastructure built for real-time interactive learning.
Data flow
Student voice and text interactions are processed through the tutoring stack, where speech tasks and educational dialog context are routed to the QwenMax and Bailian-trained model layer. Responses are returned to the learner in real time, while SLS and ARMS monitor the inference path and trigger rapid scaling when demand spikes at peak study periods.
Solution architecture
6 components · 3 layers
Compute
Elastic Compute Service (ECS)Provides the stable elastic compute base for the real-time tutoring platform.
GPU cloud serversSupply the accelerated compute used for deep-learning iteration and AI tutor inference.
QwenMaxActs as the core foundation model for tutoring, homework correction, and educational dialog understanding.
Orchestration
BailianHosts the education-vertical training flow used to build the Hanzue Teacher model stack.
Governance
Log Service SLSCaptures operational telemetry and supports rapid detection of AI-service bottlenecks.
Application Real-Time Monitoring Service (ARMS)Provides end-to-end observability and supports automatic expansion when demand spikes.
Architecture clues
Peak demand is highly time-dependent, so 30-second elastic expansion is treated as a core product requirement.
QwenMax supplies the base reasoning layer, while Bailian is used to train a domain-specific education model for the super-agent.
The system handles multimodal tutoring tasks including dialog, speech, and structured educational content extraction.
The tutoring stack combines infrastructure observability with an education-vertical model rather than using a generic chatbot directly.
Evidence from the source
Jingzhunxue uses ECS and GPU servers together with SLS and ARMS to build a stable elastic compute base and achieve 30-second scale-out plus compute self-healing. (original: 智能精准学依托阿里云 ECS 和 GPU 服务器，构建稳定弹性算力底座，通过 SLS、ARMS 监控实现 30 秒内自动扩容与“算力自愈”。)
The company deeply applies QwenMax and uses Bailian to train an education-vertical model behind Hanzue Teacher, which supports tutoring, homework correction, and question-bank extraction. (original: 智能精准学是首批深度应用通义千问 QwenMax 大模型的教育企业... 在通义大模型的基础和百炼平台能力训练教育垂直大模型。)
The tutoring product must finish real-time audio-video dialog, ASR, TTS, and formula recognition in 1.2 to 1.8 seconds to match human conversational expectations. (original: 处理实时音视频对话、语音转文本（ASR）与文本转语音（TTS）转换... 且需在 1.2-1.8 秒内完成响应。)
Traffic surges by several multiples in the evening and across weekends, which means compute must expand instantly even when GPU resources are tight. (original: 晚上 6-10 点流量较白天骤增数倍，周末全天持续高峰，流量峰值可达日常的五六倍。)
Review how this reference was scored and trace each normalized claim back to its documented source.
01
Scoring methodology
How scores are calculated
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Customer reference value100/100
Problem clarity20/20
Measures whether the customer's operational problem is specific enough to compare.
Implementation depth25/25
Rewards named components, data flow, integrations, and architecture detail.
Outcome specificity20/20
Measures whether the reported benefit is concrete rather than a generic claim.
Quantified impact25/25
5 metrics describe measurable impact.
Deployment scale10/10
Rewards evidence of a live or scaled deployment and meaningful integration scope.
Evidence confidence95%
Source authority20/25
The main evidence is an official vendor-published customer story.
Claim traceability25/25
12 of 12 key fields link to a named source.
Direct evidence20/20
4 direct source quotes retained.
Date provenance10/10
The publication date comes from the source page or publisher.
With your OK, we use Google Analytics cookies to see which pages and searches are useful. No ads, no profiling, and
nothing is loaded until you accept. You can change this any time under “Cookie settings” in the footer.
