JD Retail Technology 2023 Highlights: AI‑Driven Supply Chain, Large Language Models, Edge AI, Data Security, and 3D Modeling Innovations

Over the past year JD Retail’s technology team focused on open ecosystem building, low‑price strategies, and rapid AI adoption, showcasing 11 representative technical achievements.

Supply‑chain AI: Introduced an end‑to‑end inventory management system based on deep neural networks that directly outputs optimal replenishment decisions, eliminating the traditional two‑step forecast‑then‑optimize workflow, and launched explainable AI techniques that make model recommendations transparent, improving decision accuracy and adoption.

Advertising technology: Applied privacy‑preserving pre‑training and multi‑party secure computation to enhance user‑item matching, built a massive ranking model with hundreds of billions of parameters, and achieved a breakthrough in GPU scheduling that doubled throughput compared to TensorFlow.

Large language model framework: Developed a unified ReAct/SFT/RAG platform enabling efficient fine‑tuning, deployment, and retrieval‑augmented generation for enterprise use cases, achieving >40% training speedup and supporting models beyond 70 B parameters.

Search‑recommendation upgrades: Implemented a pre‑training‑then‑fine‑tuning paradigm for session‑level user behavior prediction, introduced a model‑plus‑policy approach for personalized main‑image distribution, and leveraged multi‑loss joint supervision for large‑scale image retrieval, significantly improving relevance and latency.

Merchant system transformation: Integrated native, Flutter, and Taro technologies for seamless cross‑platform functionality, employed OCR, RPA, and large‑model semantic understanding for automated merchant onboarding, and introduced dynamic main‑image overlay generation, cutting onboarding time to minutes.

AIGC creative generation: Built end‑to‑end pipelines for image, copy, and multimodal ad creation using category generators, personalized style generators, and a Plan‑&‑Render framework, resulting in high‑quality, scalable advertising assets and multiple top‑conference publications.

Edge AI (端智能): Delivered a 1.9 MB highly quantized inference engine supporting Android, iOS, and HarmonyOS, achieving >99% success rate over billions of daily inferences and enabling cross‑device model distribution via a unified Python‑VM platform.

Data security house: Created a “safe‑house” system combining data sandboxing, federated learning, and multi‑party computation, providing “usable but invisible” data access while integrating with Hive, Spark, and TEE‑based memory encryption.

Data asset acceleration: Implemented proactive metadata‑driven materialization, dynamic cost‑aware storage decisions, and a visual grammar‑based analytics layer, reducing dashboard build time from days to hours and scaling daily metric consumption to tens of millions.

Macro‑map system: Launched an LBS‑driven grid‑based operation platform for instant retail, standardizing data pipelines, front‑end interfaces, and high‑performance query handling to optimize supply‑demand matching across channels.

3D modeling pipeline: Designed a full‑stack solution—from a guided capture app to a proprietary .jdv format and space‑coding algorithms—that compresses 400 MB raw captures to <10 MB while delivering photo‑realistic, interactive 3D product views, now deployed on hundreds of SKUs.

These innovations collectively illustrate JD Retail’s commitment to advancing AI, data security, and immersive technologies, positioning the company at the forefront of e‑commerce research and practice.