Dazhihui and Alibaba Cloud rebuild market-data software as an AI-native app
A case study in TMTPost's Yunqi Strategy Reference describes how Dazhihui, a Chinese market-data software vendor founded in 2000, worked with Alibaba Cloud to rebuild its products as AI-native apps; talks ran through the summer of 2024 and the project started that October. The work covered four areas: reworking 20-plus years of market, financial and news data so it is unambiguous for agents rather than humans; an agent framework on Alibaba Cloud's FinXScope 2.0 with orchestration, tool-execution and security-sandbox layers, plus a six-category Skill system for quotes, stock screening, proprietary indicators, research notes, trading assistance and enterprise data queries; a four-layer memory system; and human-plus-model evaluation. The article says intent recognition that took about 10 seconds with a third-party full-size model fell to 2-3 seconds with a smaller model and to milliseconds after intent classification was trained into a 4B model, with different sizes of Qwen models matched to different scenarios and GPU, T-Head PPU and elastic compute plus on-site FDE delivery supporting the build.
It shows how a long-established financial software vendor combined data governance, agent engineering and small-model tuning in production, a reference point for similar industry AI rebuilds.
本文摘自《云栖战略参考》，这本刊物由阿里云与钛媒体联合策划。目的是为了把各个行业先行者的技术探索、业务实践呈现出来，与思考同样问题的“数智先行者”共同探讨、碰撞，希望这些内容能让你有所启发。
金融业是数字化、智能化的先行者，也是大模型技术落地的绝佳领域。从早期的电子交易系统到如今的智能投顾，技术每一次迭代都重塑着这个行业的运行逻辑。
当大模型的热潮席卷各行各业时，一个更深层的命题摆在传统金融软件商面前，如何将多年积累的数据资产与业务理解，转化为真正可用的智能能力，而非简单地接入一个 AI 对话框。金融数据是一个极其丰富、复杂的数据资产池，数据含义、口径、来源、时效性，每一个信息的偏差都可能让 AI 输出错误的回答，从而误导用户决策，这不仅是简单的财务损失，而是彻底摧毁用户的信任。所以，金融行业的 AI 改造，并不是接入一个大模型接口就能交差，而是要从底层数据治理到上层交互逻辑，完成一场由内而外的重构。