Lin Ma
Technical Director at SenseTime
Malin is a Technical Director at SenseTime and currently serves as the frontend technical lead for the Raccoon (Xiaohuanxiong) product line. He has long focused on product engineering practices for AI Native Applications. As one of the core technical leaders, he led the development of the Office Raccoon AI PPT online editing system and built an HTML-to-editable-PPT rendering engine. His work provides a scalable and reliable closed-loop interaction framework for the engineering implementation of generative AI products.
Topic
From Generation to Editability: Engineering Practices for an AI PPT Online Editing System
This talk will focus on the dual-path technical evolution of Lingdong AI PPT, illustrating how we transitioned from “large-model content generation” to a “scalable, editable online PPT system.” The first path relies on HTML-structured output: using a self-developed rendering engine to convert HTML into an editable PPT model, enabling structured content and fine-grained editing capabilities. The second path leverages raw image generation models to produce visual layouts, which are then parsed into a JSON structure for front-end re-rendering, delivering a visually prioritized generation experience. The presentation will break down the differences in system design between the two technical paths, strategies for generation stability, mechanisms for multi-round editing consistency, and how to build a full “generate → parse → structure → render → export” closed loop under complex generation scenarios. The discussion emphasizes real engineering practices and architectural trade-offs, exploring how AI-native applications can evolve from demos into sustainable product systems. Outline: * Product positioning and technical path selection for AI PPT * Path 1: HTML → editable model structured conversion design * Path 2: Raw image model → JSON rendering system implementation * Comparison of advantages, disadvantages, and scenario adaptation for the two generation paths * Multi-round generation and editing consistency control * Rendering performance and front-end data model design * How generative systems build a stable execution loop Audience Takeaways: Understand the design principles behind the “dual-path architecture” for generative applications Master engineering approaches for transforming unstructured generation outputs into editable systems Gain practical experience in ensuring stability and consistency across multi-turn interactions Learn the trade-off strategies between visual generation and structured generation in product design
