Technical Case Study of JD.com 6.18 Promotion: Intelligent Text Recognition and Order Trajectory Display

The article describes two key technical solutions implemented for JD.com's 6.18 promotion. The first is an intelligent text recognition service that automatically extracts name, phone number, region and detailed address from user‑pasted text, addressing common issues such as missing fields and misspelled administrative areas.

To achieve high accuracy, a Bi‑LSTM‑CRF deep‑learning model was designed, leveraging large‑scale data for tokenization and labeling, and using statistical sampling to ensure convergence during training.

Performance optimization for this service involved load testing at five times the expected traffic, capacity planning to determine per‑instance TPS/QPS, cache hot‑cold separation, group isolation, hot‑loading mechanisms, and SLA degradation drills, ultimately guaranteeing 99.99% system availability.

The second solution focuses on order‑trajectory display, enabling users to view real‑time package locations after payment. Raw GPS points are processed through denoising, aggregation, and thinning, then fitted to a road network to produce a smooth visual path.

Multiple algorithms were combined—including grid‑distance, K‑D tree, convex‑hull, and an optimized Douglas‑Peucker method—for point reduction and fitting. Special handling of bridges and narrow turns was achieved by simplifying the DE‑9IM topology model to two‑point segment checks, reducing computational complexity and delivering a performant, visually clean trajectory on the order detail page.