Prototype
AI Browser
负责浏览器 AI 搜索、智能问数、模型微调与部署，服务高校近万名师生并沉淀评测与推理优化链路。
系统快照
架构、信号与证据会先于正文呈现。- Users
- 10000+
- AI features
- 10+
- Eval
- RAGas
Problem
浏览器中的 AI 任务并不是一次问答，而是搜索、阅读、定位、点击、验证和总结的连续过程。项目服务高校近万名师生，覆盖 AI 搜索、智能问数、模型微调与私有化部署。
Constraints
- 网页结构不稳定，DOM、视觉布局和内容摘要都可能变化。
- 浏览动作需要可回放，方便定位失败原因。
- 推理延迟会直接影响用户体验。
System Design
原型把页面观察、检索摘要、动作规划和执行验证拆成独立模块。模型侧选用 Qwen2.5，基于校内新闻语料完成清洗与预处理，采用 LoRA + DeepSpeed 高效分布式微调，并基于 VLLM 完成推理优化和私有化部署。
评测侧负责 benchmark、并发压测与 RAG 评估，持续监控 latency、throughput、TTFT，并结合 Locust、RAGas、LLM-as-a-Judge 优化效果。
Key Decisions
浏览任务不能只依赖页面全文摘要。更稳妥的方式是结合 DOM 结构、可点击元素、页面截图和任务上下文，让模型获得足够但不过量的信息。
Impact
- 支撑高校近万名师生使用，落地十余项 AI 功能。
- 沉淀从模型选型、训练、部署到评估的全链路工程化能力。
- 建立可量化性能评估体系，为后续模型迭代和容量规划提供依据。
Reflection
后续更值得投入的是任务评测、动作约束和失败归因，而不是继续堆更大的提示词。
Problem
AI tasks inside a browser are not a single question-answer exchange. They are continuous chains of search, reading, locating, clicking, verification, and summarization.
The product served nearly 10,000 campus users and needed AI search, intelligent Q&A, fine-tuning, and private deployment to work as a usable system.
The model needs to observe page state and also know when to stop.
Constraints
Web pages are unstable: DOM structure, visual layout, and content summaries can all change.
Browser actions need to be replayable so failures can be diagnosed.
Inference latency directly affects the user experience.
System Design
The prototype splits page observation, retrieval summaries, action planning, and execution verification into separate modules.
Qwen2.5 was selected and fine-tuned with LoRA and DeepSpeed over cleaned campus-news data, then deployed through VLLM.
Evaluation monitored latency, throughput, TTFT, and answer quality through Locust, RAGas, and LLM-as-a-Judge.
Key Decisions
Browser tasks should not rely only on full-page summaries.
A stronger approach combines DOM structure, clickable elements, screenshots, and task context so the model sees enough information without being overloaded.
Impact
The project supported nearly 10,000 students and teachers, shipped 10+ AI features, and produced an end-to-end workflow from model selection to deployment and evaluation.
Reflection
Future work should focus on task evaluation, action constraints, and failure attribution rather than simply expanding prompts.