案例研究
企业级大模型与智能应用平台建设
构建统一的企业级智能数据与 AI 平台，把数据管理、模型服务、多智能体调度与知识检索整合成可复用底座。
系统快照
架构、信号与证据会先于正文呈现。- Efficiency
- 10x
- Intent accuracy
- 95%+
- Serving
- VLLM
Problem
企业 AI 应用容易分散在数据系统、模型服务、业务 CRUD、知识库和临时脚本里。真正的问题不是再做一个助手，而是建设一个能够连接企业数据、模型能力和业务流程的统一底座。
Constraints
- 结构化数据和非结构化数据需要被统一检索和调用。
- 模型能力需要可部署、可评测、可路由，不能每个业务功能都重新做一套集成。
- 高频意图识别不能完全依赖大模型调用，需要更低延迟、更稳定的轻量模型。
System Design
平台按数据层、模型层和应用层拆分。数据库、向量库、模型服务、Agent 编排和业务工作流通过受控接口连接，避免把 AI 能力做成孤立功能。
多智能体系统接入本地库、企业知识库和外部资源，完成自动检索、推理分析和结构化报告生成。模型服务侧支持医疗大模型 CPT、SFT、RLHF 与 DRPO 对齐流程，并基于 VLLM 部署上线。
Key Decisions
先标准化模型服务与业务接入，再扩展更多 AI 功能。高频意图识别由轻量 BERT 模型承担，准确率达到 95%+，推理速度约为大模型调用的 1/10；复杂推理和生成再交给 LLM。
Impact
- 报告生成效率提升近 10x，从半月级人工流程缩短到约 30 分钟。
- 沉淀统一模型服务、RAG 与多智能体框架，成为后续业务系统复用的长期基础设施。
- 将数据、模型、知识能力整合为标准化输出，支持多个业务部门快速接入。
Reflection
下一步应该把评测、可观测性和权限感知编排做成平台级能力，而不是在每个项目里重复补齐。
Problem
Enterprise AI work was scattered across data systems, model services, business CRUD workflows, and isolated knowledge bases.
The real problem was not a single assistant. It was building a shared foundation that could connect enterprise data, model capability, and business workflows.
Constraints
Structured and unstructured data needed to be retrieved through a consistent interface.
Model capabilities had to be deployed, evaluated, and routed without turning every business feature into a one-off integration.
High-frequency intent recognition needed lower latency than calling a full LLM for every request.
System Design
The platform separates data, model, and application layers. Databases, vector stores, model services, agents, and business workflows are connected through controlled interfaces.
Multi-agent workflows combine local databases, enterprise knowledge bases, and external resources to produce structured reports and reusable business outputs.
Key Decisions
The platform should standardize model serving and business integration before expanding more isolated AI features.
A lightweight BERT intent model handles high-frequency routing, while LLMs focus on heavier reasoning and generation tasks.
Impact
The platform improved report generation by nearly 10x, reducing a half-month manual workflow to roughly 30 minutes.
It also produced reusable model serving, RAG, and multi-agent components for longer-term enterprise AI infrastructure.
Reflection
The next step is to make evaluation, observability, and permission-aware orchestration first-class platform capabilities instead of project-specific add-ons.