Looking for a job

Michael Li

@michaelli

Message

Senior machine learning engineer building large-scale recommendation, ranking, and LLM systems for real users.

United States

Message

What I'm looking for

I’m looking to build end-to-end ML systems that perform for real users at real scale—strong relevance for recommendations and LLMs, with low latency, controlled cost, and close collaboration with product and safety teams.

I build machine learning systems that hold up under real users, real scale, and real-world constraints. At Snap Inc., I’m known for end-to-end ownership—turning ambiguous problems into production systems that improve reliability.

I led design and deployment of large-scale recommendation and ranking systems for Spotlight and Ads on Snap’s ML platform (Bento), addressing cold-start and relevance challenges across millions of candidates. I built and optimized two-tower retrieval for candidate generation, reducing retrieval latency by 30% while improving engagement metrics by 5% at billion-scale inference throughput.

To keep models grounded in behavior, I developed real-time and batch feature pipelines using Prism (Spark) and Robusta, solving feature freshness and enabling sub-minute updates for behavioral signals. I also drove integration of LLM-based systems using RAG in My AI, reducing relevance, hallucination, and reliability issues by grounding generation on location-aware and user-specific data.

Earlier, at Best Buy, I designed semantic product search models using in-house BERT (B3) and two-tower architecture, resolving long-tail query gaps and improving relevance. I also built personalized ranking models and store-level demand forecasting systems, and prior to that developed Bi-LSTM models for Alexa NLU at Amazon—bringing a consistent focus on measurable impact.

Experience

Work history, roles, and key accomplishments

Current

Senior Machine Learning Engineer

Current

Snap Inc.

Feb 2022 - Present (4 years 3 months)

Led design and deployment of large-scale recommendation and ranking systems for Spotlight and Ads on Snap’s Bento ML platform, improving engagement metrics by 5% at billion-scale inference throughput. Built and optimized two-tower retrieval models and LLM-driven RAG features for My AI, reducing retrieval latency by 30% and improving response relevance and grounding.

Recommendation Systems Ranking Systems Two Tower Retrieval Embedding Training Negative Sampling Robusta RAG A B Testing Lang Chain

Senior Machine Learning Engineer

Best Buy

Dec 2017 - Oct 2021 (3 years 10 months)

Designed semantic product search models using in-house BERT and two-tower architectures, improving long-tail query relevance on large-scale e-commerce traffic. Built personalized ranking and store-level demand forecasting systems that increased engagement by 60% and supported ship-from-store and curbside pickup through better inventory allocation decisions.

Semantic Search BERT Personalized Ranking Apache Spark BigQuery Databricks MLFlow

Data Scientist

Amazon

Mar 2017 - Dec 2017 (9 months)

Developed Bi-LSTM models for Alexa NLU tasks including intent classification and named entity recognition, improving language understanding for new-language expansion. Built sequence-based demand forecasting models to improve inventory planning and reduce stockouts.

Bi LSTM Named Entity Recognition TensorFlow Sagemaker Docker demand planning Sequence Modeling

Education

Degrees, certifications, and relevant coursework

University of California, Berkeley

Master of Science, Statistics

2014 - 2016

Grade: 3.7

Activities and societies: Research Assistant

Master of Science in Statistics at the University of California, Berkeley while serving as a Research Assistant. Worked on CNN models for object detection and multi-label classification using TensorFlow.

University of California, Berkeley

Bachelor of Science, Statistics

2010 - 2014

Grade: 3.8

Activities and societies: Research assistant

Bachelor of Science in Statistics at the University of California, Berkeley. Developed image preprocessing and augmentation pipelines and evaluated models using metrics such as mAP and F1.