JD Retail’s 2024 Tech Innovations: AI, Supply Chain, and Immersive Shopping

Each retrospective aims to move forward better. In 2024, JD Retail Technology supported business growth while exploring large‑model applications, intelligent supply chain, end‑technology, XR experiences, and more.

JD APP Major Redesign for a "Better and Cheaper" Experience

In the second half of 2024, JD APP launched a phased upgrade focusing on user‑mindset building. New homepage sections, a refreshed JD Life self‑operated service area, a more relevant message center, higher‑efficiency product details, smarter search recommendations, and large‑model features help users quickly find desired goods and enjoy a happier shopping journey.

The redesign introduces four core mindsets—order lookup, rights, services, and interactive play—enhances the message center, clarifies key purchase information, highlights low price, selling points, and reviews, and upgrades the Plus, flash‑delivery, and marketing channels to deliver a "better and cheaper" and smooth experience.

Data‑Driven Inventory Selection and Allocation Algorithm

JD’s supply‑chain capability ranked among the world’s best, and in 2024 JD became the only Chinese retailer in Gartner’s 2024 Global Supply Chain Top 25. The team applied a data‑driven inventory selection and allocation algorithm to balance stock across regional distribution centers (RDCs) and front‑line distribution centers (FDCs), reducing out‑of‑stock rates and fulfillment costs while supporting tens of millions of orders with sub‑24‑hour delivery for over 90% of self‑operated orders.

The algorithm analyzes historical co‑purchase patterns and demand forecasts to maximize order fulfillment rate, and an end‑to‑end allocation model outputs the highest‑stock‑availability, lowest‑cost plan with strong interpretability, delivering faster fulfillment for over ten million daily orders.

JingDian AIGC Platform for Automated Marketing Content

The JingDian AIGC platform enables merchants to generate high‑quality product images, copy, and short videos with a single click. It leverages a self‑developed text‑to‑image foundation, ReferenceNet and ControlNet image models, and a multimodal product‑understanding model combined with RAG and large language models to produce accurate, market‑aligned marketing copy. Reinforcement learning continuously refines the models based on user feedback.

To date, the platform serves over 350,000 merchants, handles more than 10 million AI calls per day, and has won the “InfoQ 2024 China Technology Power – Best AI Practice” award.

Low‑Code National‑Subsidy Platform for "Trade‑In‑New" Initiatives

In response to nationwide "trade‑in‑new" subsidies, JD built a reusable low‑code national‑subsidy qualification platform that can be configured for different cities within minutes, enabling rapid rollout of subsidy services. Integrated AI‑driven risk modeling automatically blocks over 99% of scalper accounts, keeping the risk user ratio below 0.5%.

The platform now supports multiple subsidy forms—order‑level discounts, trade‑in discounts, payment discounts, and coupons—delivering a "better and cheaper" shopping experience across JD’s multiple channels.

Data Lake Architecture Innovation

JD adopted Apache Hudi for its data lake, focusing on I/O performance, feature richness, and ecosystem integration. The team built multi‑modal I/O capabilities, lock‑free concurrent writes, and linear scalability, contributing over 65 PRs to the open‑source community. The lake now stores 160 PB of data, achieving minute‑level latency (T+1 to minutes) and reducing storage‑compute costs by more than 12 million RMB annually.

JD Merchant Intelligent Assistant

The assistant provides 24/7 AI‑driven support for merchants, handling queries, order management, customer service, data analysis, sales forecasting, pricing, and keyword recommendation. Built on a large‑language‑model multi‑agent system with dynamic planning, it achieves over 90% decision accuracy and sub‑second response times, dramatically improving merchant efficiency.

Taro on Harmony: One Codebase for Multiple Platforms

Taro provides an open‑source cross‑framework solution for mini‑programs, H5, and React Native. The Harmony extension lets developers write React DSL code that compiles to native Harmony apps, reducing development barriers and enabling rapid, low‑cost migration of JD’s Android/iOS app to the Harmony ecosystem.

Ad Creative Generation and Personalised Recommendation

The advertising team introduced a multimodal reliability feedback model to improve generated image usability and released the first human‑annotated RF1M dataset for ad image generation. They also split creative selection into element and composition stages using multimodal LLMs, boosting cold‑start performance and achieving precise, large‑scale personalized ad recommendations.

JD Vision & “LiYing” 3D Plan for Immersive Shopping

JD launched JD Vision, the first transaction‑capable Apple Vision Pro app in China, leveraging spatial computing, eye‑tracking, and hand‑gesture interaction to place 1:1 product models in users’ real environments. The subsequent "LiYing" plan adds shift‑lens rendering and depth‑culling to deliver a naked‑eye 3D product showcase, creating a playful, immersive shopping experience.

AI Engine Built on Domestic Chips

To address the rise of Chinese AI chips, JD’s Nine‑Number algorithm platform created an end‑to‑end AI engine compatible with both NVIDIA GPUs and domestic NPUs. The engine offers unified APIs, model quantization, compilation optimisations, and MFU‑level performance boosts, enabling zero‑cost model training and deployment across diverse hardware.