Meituan has built DeepSkill, a Dianping-style skill platform that includes more than 370,000 skill tools.

Meituan AI is also ramping up its efforts.

Exclusive information from Dujia shows that DeepSkill, a skill platform for AI Agent developed by Meituan AI team, has included more than 370,000 Skill tools verified through real tasks. Back in early September, this figure was 130,000, which has more than doubled. Compared with common AI plugin markets, this platform has designed a set of user data feedback mechanism, aiming to address practical problems in skill selection and data security during the use of Agents.

However, there are quite a number of Skill platform products or communities, and leading tech giants such as Tencent, ByteDance and Alibaba have all made layouts in this field. What makes Meituan different?

What is DeepSkill

There is a huge number of available Skills on the market, but it is difficult to judge whether they can adapt to your specific tasks only by reading the profiles written by developers. Meanwhile, third-party Skills are uneven in quality, and users will worry about the leakage of private data that is secretly uploaded when processing sensitive content such as resumes and personal information. The entire product process of DeepSkill is designed to solve these two pain points.

It does not adopt the traditional directory browsing mode. Users do not need to filter one by one in the classification bar in advance. They only need to tell their Agent the complete task requirements directly, and the platform will automatically decompose the task demands and match suitable candidate Skills from the massive library. The platform will not only rely on the self-description of the uploader, but also scan and parse the full content of the Skill, output a complete analysis report that lists all capability boundaries, calling tools and potential risks, including data security review, which clearly marks whether private materials such as user-input resumes will be uploaded, directly eliminating the concern of information leakage.

The core highlight of the platform lies in the reviewer data feedback mechanism. Every user who actually uses the Skill can become a reviewer. The system will automatically sort out the running result of this task, and users can submit evaluations with simple confirmation, without writing long paragraphs manually.

Scores, real task cases and actual operation effects are all made public, so that other users can refer to the real feedback of predecessors in similar scenarios to make decisions. For example, in the interview preparation scenario, users can view the real records of others using this set of Skills to match resumes with JD and sort out interview question banks, to assist their own selection and reduce trial and error costs. Every real use in turn enriches the public data of the community, forming a closed loop of "use - review - reference".

The Agile Iteration of Meituan AI

From an industry perspective, Meituan's development of DeepSkill is not just about building a plugin download store. Offline businesses such as food delivery, in-store services and merchant operation themselves require a large number of AI automation capabilities. Meituan has accumulated a lot of Agent implementation experience internally, and Skill is the basic component for AI to perform complex tasks.

At present, there is a practical problem in the AI industry: various AI frameworks are iterating rapidly, but the supply of stably available Skills verified in real scenarios is insufficient. Many Skills only perform well in demonstration scenarios, and are prone to failure when implemented in actual tasks. Many similar platforms are only responsible for putting Skills on shelves for collection, lack of large-scale actual measurement and unified security verification, making it difficult for users to judge the practicability and risks of Skills.

Building DeepSkill is valuable both internally and externally. Externally, Meituan can export its practical experience in internal Agent implementation, expand the skill library relying on community crowdsourcing, and screen reliable Skills with the help of a large number of real users' usage processes. Internally, the high-quality skills precipitated by the community can also feed back Meituan's own business. If an Agent wants to be implemented on a large scale, it is very difficult for an enterprise to develop it alone, and an open community is a feasible path. With the help of user feedback, the platform converts personal usage experience into public reference information, lowering the entry threshold for developers and ordinary users.

Compared horizontally with other leading tech giants, Meituan has an overall low-key style in the AI field. ByteDance, Alibaba and Tencent will focus on promoting large models and AI products externally. Although Meituan has self-developed large models and many AI tools have been implemented in the scenarios of delivery riders and merchants, most of them prioritize serving its main business, and few highly popular AI products for the general public have been launched.

This strategy stems from its business philosophy. Meituan pays more attention to whether AI can solve real problems. New products often go through a long period of polishing before being released to the public, so does DeepSkill. It has not been marketed on a large scale, and is directly handed over to the market for testing.

This set of strategies has both advantages and disadvantages. The advantage is that the product is more pragmatic and reduces conceptual gimmicks; the shortcoming is that the outside world tends to underestimate its technical layout. DeepSkill is a new attempt by Meituan, which is no longer limited to serving internal businesses, but tries to export Agent ecological capabilities externally. What effect it can achieve in the future remains to be seen. #Meituan #Meituan AI #DeepSkill #Meituan Ramping up AI #Dianping-style Skill Platform #Dianping-style AI Platform#