Wenshan Wu Email & Phone Number

Beijing, China

Beijing, China

working on AI for language learning, LLM for task automation in MSRA.Mispronunciation diagnosis retrieval system Responsible for the launch of the brand-new feature, including proof of concept, feature design and development. Implemented the one-stop framework as a data-driven solution which supports unstructured data search, pattern mining and indexing, and content generation as an AI + HI solution. Built a pure cloud microservice solution from scratch based on K8S infrastructure with API-gateway, service mesh and loggings, meeting the goal of observability (both service health and business metrics), scalability (end-to-end scale metrics like QPS, response latency), optimizing debugging and issue tracing. Built a human workflow platform for content generation in AI + HI scenario, which is configurable and extendable, relieving redundant efforts from backend when applied to other tasks.Microsoft AimWriting for writing polishing Responsible for backend solution, including feature developing, service migration and data warehouse for user analyzing.

Shenzhen, Guandong

responsible for Wechat game platform development and security strategy development in WeChat Team.WeChat Minigame Security platform Responsible for security challenge in WeChat mini-game scenario. Developed a content perceiving system by image and text clustering cross all games, thus get control of popular content in our platform within one hour, helping to decrease violation rate from 5% to 0.5%. Developed a system for data analysis, real-time monitoring alert based on druid. Provide one-stop service including abnormal perceiving, related data analysis and corresponding strategy preview.WeChat Esports platform Responsible for modules of Esports competition and video feeds. Developed a deep learning based real-time recognition solution for heroes in the game "Honor of Kings" and numbers in game "PUBG Mobile". The real-time video tags improved click-rate by 30%. Implement automatic data collecting and synthesizing tools for continuous operation, collecting millions of semi-labeled images without human labeling by taking advantage of crawling and image processing algorithms, cutting cost of 100 human's labor per day, and achieving over 95% accuracy of those models.