Zihao Feng

I’m a Senior Machine Learning Engineer at Snap building large-scale representation learning and generative retrieval systems. My work spans GNNs and Transformer-based modeling, along with the distributed infrastructure required to train and deploy models at scale. I’m increasingly focused on generative modeling and post-training techniques.

At Snap, I led social graph representation learning using GNNs and Transformer encoders, improving friending quality and large-scale retrieval performance. I explored generative retrieval via semantic tokenization (Semantic IDs) and next-token approaches, and I built production LLM pipelines for safe user recommendation with curated high-quality human-labeled datasets and optimized precision-recall tradeoffs. I also investigated preference optimization (GRPO-style) to align open-weight LLMs with policy objectives, and I rebuilt distributed graph training infrastructure to make training 3x cheaper and inference 20x cheaper—contributing to large-scale open-source GNN systems.

Previously at Pinterest, I helped build the company’s first neural retrieval models for ads, evolving from heuristic interest matching to GraphSAGE-based user representations and a backend two-tower deep retrieval architecture. I migrated major recommendation surfaces to NN-based retrieval, driving substantial improvements in engagement quality and monetization efficiency. I also worked on an EfficientNet-based solution to remove NSFW content from Pins and improved comment sentiment analysis by combining image content with text using OpenAI CLIP.