Risk Management
Shuyao Technology needed to replace a brittle, manually maintained financial risk rule engine. Its decision process spanned 516 nodes and 878 evaluation indicators, with tightly coupled logic, redundant checks, and slow policy updates that could not keep pace with changing fraud patterns. The company wanted a model that could understand the full decision flow, discover latent rules, and keep risk decisions explainable enough for regulated financial use.
Reported impact
Shuyao raised critical rule-matching accuracy to 86.83%, reached 98.51% similarity between model reasoning and the original decision path, improved explainability for financial audit and compliance needs, and created a reusable long-context training pattern that led the customer to expand its purchase of AI services and compute capacity.
Shuyao Technology and Alibaba Cloud built a financial risk-decision training pipeline around Platform for AI (PAI), Qwen3, and AI Stack. The team first converted BPMN decision flows into structured corpora that captured node topology, rule-package logic, and end-to-end decision paths, then used a staged recipe of continued pretraining, cold start, supervised fine-tuning, and reinforcement learning to turn the base model into a risk specialist. Alibaba also optimized the stack for long-context training by compressing 501K-token source flows into usable 57K-token samples, removing repetitive rule IDs, and applying DeepSpeed ZeRO-3, Liger Kernel, AutoTP, and YARN dynamic RoPE so 64K-context training and 40K+ token reasoning could run without truncating critical cases. The resulting model plugs into Shuyao's risk-control workflow to propose and explain decisions instead of relying only on hand-written rules.
Shuyao converts BPMN risk workflows and rule packages into structured training corpora, feeds them into PAI and AI Stack for continued pretraining, cold-start tuning, supervised fine-tuning, and reinforcement learning on Qwen3, and then uses the resulting model to analyze complex financial decision flows, recommend rules, and explain outcomes for lending risk operations.
After two reinforcement-learning iterations, critical rule-matching accuracy jumped to 86.83% and reasoning-step similarity reached 98.51% (original: 经过两轮精细迭代优化（RL 训练总耗时 113 小时），模型的关键规则匹配准确率从初始的不足 30% 跃升至 86.83%，风控推理步骤相似度（LCS）达 98.51%).
At the same cost, the project increased context support by 3.2x and produced a reusable best practice for long-text training in financial risk control (original: 在 APG 利用上实现突破——同等成本下上下文支持能力提升 3.2 倍，训练效率与显存优化成效显著).
The core objective was to let a large model deeply understand complex risk-control logic, automatically mine latent rules, and continuously optimize strategy loops (original: 核心诉求明确：让大模型深度理解复杂的风控决策逻辑，自动挖掘潜在规则，并实现策略的持续优化闭环).
The optimization stack reduced memory usage by 65% and enabled 64K-context training without truncating 40K+ samples (original: 引入 DeepSpeed ZeRO-3 + Liger Kernel + AutoTP 显存优化技术，实现显存占用降低 65%，支持 64K 上下文训练，彻底消除此前对 40K+ 样本的截断问题).
The risk-control decision process contained 516 nodes and 878 evaluation indicators, with tightly coupled logic and redundant checks (original: 其风控决策流程极为复杂，包含 516 个节点、878 项评估指标，规则之间高度耦合，决策流中普遍存在无规律符号、冗余逻辑与重复判断等问题).
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
Alibaba Cloud Sample Archived · Jul 21, 2026 EaseMate EaseMate's AI image and video generation required massive computational power with low-latency delivery. Content safety became paramount — with millions of AI-generated assets monthly, the company needed robust… Alibaba Cloud Model Studio · Qwen Alibaba Cloud Sample Archived · Jul 21, 2026 Go8 Technology Inc. As Go8 Technology expanded its portfolio of digital solutions, it required a scalable, reliable, and cost-efficient cloud platform to support growing customer demand and increasingly sophisticated workloads. The company… Reduced application latency from 100–250ms to 10–50ms through deployment on local Philippine infrastructure Application Load Balancer (ALB) · Container Service for Kubernetes (ACK) · Content Delivery Network (CDN) Alibaba Cloud Sample Archived · Jun 9, 2026 A2E As a platform aggregating 20+ AI models for video generation, A2E.AI faces significant technical challenges. Running multiple inference models — video synthesis, face swapping, voice cloning, lip-sync, and image… Alibaba Cloud Model Studio · Qwen 