JD Retail Technology 2023: AI‑Driven Supply Chain, Advertising, Edge AI, Data Security, and 3D Modeling Innovations

Over the past year, JD Retail’s technology team focused on open ecosystem construction and low‑price strategies, delivering billions in subsidies, optimizing traffic allocation, simplifying merchant onboarding, and integrating large models into numerous business scenarios to enhance efficiency and capability.

Supply‑Chain Innovation Technology earned top industry awards by introducing end‑to‑end inventory management based on deep neural networks, which directly outputs optimal replenishment suggestions, and explainable AI that makes forecasting transparent and improves adoption.

The resulting automated replenishment system achieved over 85% automation, 95% in‑stock rate, and reduced inventory turnover to ~30 days across more than 10 million SKUs.

Advertising AI introduced privacy‑preserving pre‑training, large‑scale sequence modeling, and a 100‑billion‑parameter ranking model, along with dynamic GPU scheduling that doubled throughput compared to TensorFlow.

Large‑Language‑Model (LLM) Framework combined ReAct, SFT, and RAG techniques to enable efficient fine‑tuning, deployment, and application, supporting 70B+ models with 40%+ training speedup and integrating vector database Vearch for fast retrieval.

These LLM capabilities have been applied to knowledge Q&A, user growth, sentiment mining, and data analysis.

Search‑Recommendation Upgrade introduced a pre‑train‑fine‑tune paradigm for session‑level prediction, model‑plus‑policy collaboration for main‑image personalization, and a Re‑ID based image search pipeline that dramatically improved feature discrimination.

Merchant System Deep Refactor unified native, Flutter, and Taro technologies, applied OCR, RPA, and LLM‑driven semantic understanding for fast onboarding, and launched features such as dynamic main‑image overlay for marketing.

Edge AI (Device Intelligence) delivered a 1.9 MB on‑device inference engine, achieving >99% success rate for billions of daily inferences across Android, iOS, and HarmonyOS, and integrated a cloud‑side Python VM for unified code execution.

Data Security House built a "usable but invisible" platform combining data sandbox, federated learning, and multi‑party secure computation, integrating TEE‑based memory encryption and fine‑grained access control to protect privacy while enabling data sharing.

Data Asset Upgrade introduced proactive metadata‑driven materialization and cost‑aware storage policies, dramatically improving query efficiency and reducing storage‑compute costs.

Macro‑Map System provides LBS‑based grid operations for instant retail, linking supply and demand across B2C and O2O channels to improve multi‑channel efficiency.

Low‑Cost High‑Quality 3D Modeling delivers photo‑realistic 3D product displays using a custom .jdv format, NeRF‑based reconstruction, and a lightweight client decoder, achieving PSNR > 40 dB with compression from 400 MB to <10 MB.

These innovations collectively illustrate JD Retail’s commitment to advancing AI, data security, edge computing, and immersive 3D technologies to drive e‑commerce transformation.