JD Retail Technology 2024 Innovations: AI-Driven Platforms, Data Lake, Cross‑Platform Development, and Intelligent Supply Chain

JD APP completed a phased major redesign in the second half of 2024, introducing a clearer home page, enhanced message center, smarter product detail pages, and AI‑driven search and recommendation features to deliver a "better, cheaper" user experience.

The company deployed a data‑driven inventory selection and allocation algorithm across its regional distribution centers (RDCs) and front‑line distribution centers (FDCs), reducing out‑of‑stock rates, cutting fulfillment costs, and enabling over ten million orders to be delivered within 24 hours.

JD’s self‑built AIGC platform "JingDian" provides merchants with high‑quality product images, copy, and short videos using proprietary text‑to‑image models, multimodal product understanding, and reinforcement‑learning‑enhanced generation, serving more than 350 000 000 daily AI calls for over 350 000 merchants.

A low‑code national‑subsidy qualification platform was rapidly launched to support the 2024 "old‑for‑new" subsidy campaign, leveraging reusable integration models and AI‑enhanced risk control to block over 99 % of scalper activity.

JD’s data lake, built on Apache Hudi, now stores 160 PB of data, achieving minute‑level latency (T+1 to real‑time) and reducing annual storage‑compute costs by over 12 million USD, while supporting real‑time traffic‑driven operations during major sales events.

The Merchant Intelligent Assistant, powered by a multi‑agent LLM system, offers 24/7 conversational support for merchants, achieving over 90 % decision accuracy and sub‑second response times across the full e‑commerce workflow.

Taro on Harmony enables a single codebase to run on HarmonyOS, mini‑programs, H5, and React Native, providing near‑native rendering performance and lowering development barriers for cross‑platform applications.

JD’s advertising team introduced a multimodal feedback model and the industry‑first RF1M dataset to improve AIGC image usability, and a two‑stage creative selection pipeline that boosts high‑quality ad generation and personalized recommendation, with results published at AAAI, ECCV, and IJCV.

The "JD.Vision" and "LiYing" projects deliver immersive AR/VR and naked‑eye 3D shopping experiences, using spatial computing, eye‑tracking, and depth‑culling technologies to let users interact with 1:1 product replicas in real space.

Finally, JD’s AI engine, compatible with both NVIDIA GPUs and domestic NPUs, provides a unified training and inference stack, enabling zero‑cost model deployment, performance gains through MFU optimization, quantization, and compilation, and supporting a range of retail AI applications.