”As an Internet logistics enterprise, Huolala has always been committed to using innovative technologies to seamlessly connect communities with simplified delivery services. Empowering users and partners through innovative logistics solutions will contribute to the digitalization process and promote the enhancement of social digital inclusiveness. For our overseas business, we leverage Amazon Web Services' capabilities and services in generative AI to explore cutting-edge technologies, optimize business operations, and integrate generative AI into our business models

——Customer reviews

summary

Huolala is a technology company specializing in the field of freight logistics. The company's core business includes intra city and cross city freight, as well as customized enterprise logistics services, errand running, car rental, and aftermarket for enterprises. The mission of the company is to create an efficient, transparent and convenient freight logistics service platform through the Internet and mobile Internet technology, provide high-quality freight services for cargo owners and truck drivers, improve logistics efficiency and reduce logistics costs. At present, the Amazon cloud technology products and solutions used by Huolala's overseas business include: Anthropic Claude 3 Haiku in Amazon Bedrock, Anthropic Claude 3 Sonnet in Amazon Bedrock, Amazon API Gateway, Amazon Lambda, etc.

Project Background

Actively embrace generative AI, balance costs, benefits, and security issues in applications, and enhance the intelligent operation level of Huolala

One of the important factors for Huolala to have a high competitive advantage is its ability to quickly and efficiently transport goods through a vast network of truck drivers and sources of goods. With the vigorous growth of its business, Huolala expects to continuously introduce the latest technologies such as artificial intelligence and big data to enhance the level of operational intelligence, achieve intelligent truck scheduling, and improve capacity utilization.

Since the emergence of generative AI technology, Huolala has quickly established a dedicated project team to evaluate the scope of its capabilities and its impact on its business. It firmly believes that generative AI will reconstruct most of the company's overseas business scenarios, such as marketing and quality improvement scenarios, and has begun to actively embrace generative AI. In addition to the continuous research and promotion of the application of generative AI technology by product and algorithm teams, more and more business departments targeting overseas customers are actively paying attention to how to apply big models to solve and optimize business problems. How to apply big language models to serve business well, Huolala mainly focuses on three aspects: the cost brought by adopting generative AI technology, the benefits brought by generative AI to business, and how to ensure data security and compliance while applying generative AI.

Cost: Huolala has a wide range of overseas businesses and multiple scenarios, with a huge number of tokens. Taking quality inspection business as an example, tens of thousands of phone calls are made every day, and quality inspection includes multiple dimensions. Therefore, the monthly required number of tokens is very large. Using industry models, whether deploying or calling APIs, the cost is relatively high;

Huolala has a solid foundation of cooperation with Amazon Web Services. Previously, it relied on Amazon Web Services' extensive, deep, flexible, reliable, secure, and compliant infrastructure to quickly achieve business breakthroughs in Southeast Asia and Latin America, creating high-performance, high

Solutions

Quickly implement innovative applications of generative AI through Amazon Bedrock, utilizing multimodal recognition capabilities to shorten interactions and enhance customer experience

The overseas business department of Huolala has a strong demand for the application of generative AI technology. Currently, many businesses have explored and practiced, including dozens of applications in 14 scenarios such as HR, PMO, invitation customer service training, quality inspection training, security risk control, finance, and financial customer service Q&A. Based on Amazon Web Services, generative AI applications have been deployed or tested for overseas business.

技术亮点 | 从模型选择与评估、提升数据质量到模型微调，亚马逊云科技端到端生成式 AI 应用创新服务让快速落地

在面向海外业务的各类场景的生成式 AI 应用部署中，货拉拉还需解决模型评估和筛选、数据准备和模型微调等问题，亚马逊云科技提供了一站式端到端的生成式 AI 服务，帮助货拉拉加速在其海外业务中的生成 AI 应用落地。

模型评估和筛选：业界LLM（Large Language Model，大语言模型）众多，由于参数量、训练数据及模型架构不同，LLM 能力存在 些显著差异，需要根据具体业务需求和场景选择合适的模型。对于 LLM 的全面的基准测试和评估，需要消耗大量的人力和计算资源。

针对货拉拉的海外业务流程，亚马逊云科技 AI 应用科学家先后对 10 款模型进行测试评估，评估数据包括 1000 条线上真实数据以及 LLM 生成的数据相结合，将准确率、召回率、任务完成率作为模型选择标准。以邀约质检场景邀约话术场景为例，Claude3 Sonnet 在基准测试中准确率为0.98，精准度为 0.99，召回率为 1，F1-Score 为 0.99，横向评估中获得第 。

数据生成和数据质量提升：通过亚马逊云科技调用 LLM，根据基本业务流程和提示词来生成数据、清洗数据并提升数据质量，使用 AmazonBedrock API 共生成 3000 条数据，并过滤掉轮次过短、说话主体不明确、没有模拟生成 API 接口输出的对话，以确保数据质量满足业务需求。

模型微调：亚马逊云科技基于货拉拉的业务需求对多款开源模型进行了微调，由于当前训练数据相对较少，因此最终采用了 Lora 的方法进行训练。此外，还尝试构造更多样化的数据及人工清洗的方式，在保证数据的质量和多样性的基础上，进行全参微调，进一步提高模型泛化能力。

项目管理亮点 | 通过亚马逊云科技 Amazon Bedrock 调用 Claude 3 模型，多模态识别让货拉拉更丝滑处理问题、提升用户体验

货拉拉的海外业务处理中，接收到的信息从原来的文字版信息，越来越多地转变为各类图文信息，接收到的图文问答、保险单比价，都需要从图片里面去解析和提取信息。货拉拉通过亚马逊云科技 Amazon Bedrock 调用 Claude 3 模型，利用其优秀的多模态识别能力进行图文问答、图文推荐、比价等，海外业务实际应用效果非常好。

Results and benefits

应用敏捷上线，生成式 AI 让业务更低成本、更好覆盖客户

生成式 AI 实验时间从 3 个月缩短到 6 周，15 个海外业务已部署生成式 AI 应用

生成式 AI 在货拉拉内部已经处于爆发阶段，很多海外业务部门带着业务的应用场景来主动来寻求产品和算法团队的支持，积极推进生成式 AI 应用的落地。通过与亚马逊云科技的通力合作，货拉拉业务的各个领域，包括针对海外业务的 14 类业务场景和数十个业务已经开始尝试采用生成式 AI，其中 15 个业务的生成式 AI 应用已经上线。

利用亚马逊云科技生成式 AI 能力，货拉拉可以减少模型训练的时间，高效检验不同 LLM 与业务场景需求的匹配度，快速进行业务创新实验，将实验时间从 3 个月缩短到 6 周，将更多的时间用来提升数据的质量和精确度及优化提示词。

效率提升、成本降低，应用上线时间从 1 至 2 天缩短为 10 多分钟

以前货拉拉上线一项新应用或功能，可能需要一至两天。货拉拉通过构建悟空平台，集成业界主流 LLM，面向公司的海外业务，通过低代码或零代码方式快速搭建应用，利用模型组件化，可视化、采取拖拉拽式的操作，仅十几分钟就能把一项业务搭建起来，大大地促进业务上线效率。引入生成式 AI 不仅提升了业务效率，还为海外业务实现了成本节省，比如培训业务中引入生成式 AI 后，除了提升司机体验外，每年预计为培训师节省上千小时的培训时长，大大降低公司的培训成本。

邀约质检覆盖抽样率提升 10 倍，更高效审核对话规范性

通过大模型进行质检，可以更多地去检测发现业务运营中的问题。以前由于人力、成本的限制，质检抽样率较低。 现在，货拉拉基于亚马逊云科技的生成式 AI 能力可以将抽检率提升 10 倍，将对客服邀约人员与司机对话的规范性进行更为广泛、高效的审核，规范邀约人员话术行为，必将增强司机的体验。

未来，货拉拉将继续与亚马逊云科技加大包括图片在内的其他多模态方面应用，包括视频、音频等，提升运营效率和服务质量，构建了更敏捷、高性能、低成本的智慧货运平台，为海外客户带去更加优质的货运