Le Yu

Senior Machine Learning Engineer

TikTok

App Ads Bidding

• Delayed GMV prediction

  • Drove the end-to-end design and launch of a hybrid delayed-GMV model for Target ROAS bidding, fusing delay curves with click and real-time funnel signals to predict final GMV before feedback matured. The initial IAA launch improved advertiser value by about 7%.

• Long-horizon bid control

  • Designed and launched a multi-day MPC controller that incorporated historical spend and predicted GMV into each bidding decision to correct accumulated delivery gaps. An IAP experiment increased the share of campaigns meeting ROAS targets by about 14% and advertiser value by about 37%.

• Generative bidding

  • Built minute-level state–action–reward trajectories, RTG targets, and training and serving pipelines for sequence-based bidding. Prototyped Decision Transformer and generator–critic–value models with controller-based action bounds and safety priors.

• Ads diagnosis agent

  • Designed and launched a production App Ads diagnosis agent for ticket parsing, funnel triage, and bidding and ranking investigations. Authored its modular Skills and integrated internal tools with evidence and confidence controls to deliver traceable diagnoses while separating verified findings from hypotheses.

Machine Learning Engineer

TikTok

Lead Ads Ranking

• Federated learning framework

  • PACS-FL keeps privacy-sensitive conversion labels in a protected sandbox, exchanging intermediate tensors and gradients with cloud workers for split training. Drove Framework 2.0 for Lead Ads: a C++/Proto training interface with a unified label-feature pipeline, validated automatic cloud-sandbox graph partitioning, and single-job orchestration across GPU batch, CPU batch, and streaming stages, cutting PS storage by 70% and batch time by 90%.

• Unified conversion model

  • Drove the staged unification of trackable and privacy-restricted traffic across on- and off-platform Lead Ads into one multi-head model with shared training and serving. Traffic-conditioned adapters and a compact sparse-feature tower preserved segment-specific behavior. The broader One Model program shipped through 18 production launches across US, EU, and ROW, with individual unification launches delivering up to 18.6% advertiser-value lift

• Privacy-compliant ad retrieval

  • Launched multiple recall branches using complementary privacy-safe signals: ELAND landing-page interaction graphs for U2I matching; Memory Recall for cluster-based candidate transfer; Cohort and DP behavioral features; and PACS-FL pCVR for eCPM retrieval.

Machine Learning Engineer

TikTok

Promotion Ads Ranking

• Ads–recommendation collaboration

  • Built bidirectional sample and feature pipelines between Promote and TikTok's recommendation system, allowing promotion models to use organic engagement signals and recommendation models to learn from paid-promotion feedback.

• New ad product launches

  • Launched Promote Internal Ads, Ad4Ad, and Promote-to-Live, building their CVR models and conversion pipelines. For Promote-to-Live, replaced daily updates with streaming training and modeled delayed watch-time labels.

Machine Learning Engineer Intern

TikTok

Machine Learning Intern

E Fund Management

• Built LSTM baselines over hundreds of proprietary market features for equity time-series forecasting.

Machine Learning Intern

Microsoft

• Fine-tuned BERT to classify articles for Bing News and support personalized newsletter curation.

Machine Learning Intern

WeChat

• Developed a ResNet-based visual model for detecting unauthorized and reposted videos on WeChat Channels.

Strategy Intern

Tencent

• Analyzed global cloud-computing markets and competitors to support Tencent Cloud product planning.

Machine Learning Intern

Microsoft

• Improved an N-gram language model for first-page result reranking in Bing Search.

Part Time Assistant

Roland Berger

Master

Peking University

Electronics and Communication Engineering

Exchange

National TsingHua University

Electronic Engineering

Bachelor

Southeast University

Automation Engineering