How Edge AI Powers Alibaba’s Local Life Services: Architecture and Real‑World Wins

Edge AI Concept

Edge AI refers to running machine‑learning or deep‑learning inference directly on client devices, enabled by modern CPUs and GPUs that can handle high‑level computations, sometimes sacrificing precision for performance, thus avoiding bandwidth costs and network latency.

Background

Both cloud and edge AI aim to process large data inputs, adjust algorithms, and infer optimal decisions; edge AI brings these capabilities to the device, standardizing intelligent algorithms, collecting real‑time user behavior and device data, and feeding back to refine algorithms, operations, and product strategies.

Features

Compared with cloud‑side AI, edge AI offers low latency, data privacy protection, and reduced cloud resource consumption.

Trends

In recent years, edge AI has become a hot R&D direction for major internet companies, powering features like AI cameras, FaceID, and AR effects in short‑video apps; Alibaba’s own products such as Taobao’s photo search, AR try‑on, and product recommendation also exemplify edge AI integration.

Local Life Scenario

The local‑life market, estimated at roughly 20 trillion yuan with a digital penetration of only 10‑15%, differs from traditional e‑commerce by emphasizing locality and real‑time service, making it a prime candidate for edge AI solutions.

Business Attributes

Ele.me focuses on home‑delivery services; rapid decision‑making during peak meal times requires matching user demand with supply instantly, a problem well‑suited for edge AI.

Technical Architecture

Multiple Alibaba local‑life apps (Ele.me, Koubei, Fengniao, Merchant version, XuanYuan) share a unified edge‑AI adaptation layer built on Alibaba’s middleware and the Edge AI SDK. The architecture comprises two foundations:

Device foundation: UT telemetry, Orange feature flags, Highway high‑speed data transfer, Answer real‑time monitoring, SQL database access, Mtop network library.

Edge‑AI foundation: MNN inference engine, BehaviX user behavior data, Walle runtime environment.

An ALSCAdapter layer adds lifecycle management, automatic telemetry, and opens APIs for debugging and testing, integrating with the MNN workbench to boost algorithm and testing efficiency.

Technical Exploration

Ele.me has explored edge AI in user profiling, intelligent recommendation, and smart push.

User Features

Rich, real‑time user behavior data (store entry/exit, add‑to‑cart, browsing, etc.) are collected on‑device, forming a metadata pool that feeds both cloud and edge models, enabling rapid feature updates within ~2 seconds.

The feature extraction pipeline includes on‑device collection, graph back‑propagation, and persistent storage for reuse across models.

Intelligent Recommendation

Traditional cloud‑based recommendation suffers from minute‑level latency, unsuitable for real‑time feeds. Edge AI enables on‑device re‑ranking of unseen stores based on recent user actions, allowing multiple decision points per session and improving relevance.

Smart Push (Intelligent Reach)

Edge AI models deployed on the client identify user behavior turning points, collaborate with cloud push engines via a “race” mechanism to select optimal push configurations, resulting in higher click‑through rates and better user retention without intrusive notifications.

Other Business Cases

Projects like Fengniao’s “Blue Storm” use on‑device pre‑recognition to streamline rider photo verification, reducing manual effort and improving service quality.

Future Outlook

As mobile hardware and network infrastructure improve, edge AI’s role will expand, especially for real‑time search recommendation and operational intelligence in local‑life services. Ongoing efforts aim to solidify edge AI as a reusable infrastructure for rapid business integration.