Luke Salamone

• Trained transformer-based T5 query denoising autoencoder, increasing retrieval coverage by 36%. Optimized inference code and deployed to Kubernetes cluster.
• Directed ML intern in end-to-end development of query intent classifier. Directed data instrumentation, architectural exploration, and production search inference optimization.
• Led cross-functional team to design and integrate two-tower writing action recommendation model. Assisted in offline evaluation and production architecture deployment.
• Developed pairwise knockout selection and iterative self-play techniques for synthetic data generation, enabling rapid experimentation.

• Increased click-through-rate by 2.1% and staytime 1.4% via introduction of efficient DCNv2 architecture in search ranking, while decreasing serving costs by 10%.
• Improved AUC on cold-start document CTR prediction by 0.8% over baseline via training of pointwise multitask learning model.
• Improved music retrieval HNSW index with click-query deep retrieval system, enabling a 15% nDCG improvement and 0.55% increase in overall music add rate in music search.

• Implemented state-of-the-art conversational chatbot utilizing vector-based knowledge retrieval. Chatbot now available to over 100 million monthly users.
• Improved website search results by 37% (MRR) through optimization of ML model.

• Reduced ETL costs 90+% by building a high-volume Java library for data streaming.
• Reduced errors by 95+%, lead development of serverless app for data integrity / reporting.
• Designed data recommendation tool for internal metadata inquiries utilizing a hierarchical classification machine learning system.

Chessmate AI – Trained Vision Transformer (ViT), achieving state-of-the-art chess move prediction (55% top 1). Fine-tuned for personalization via Direct Preference Optimization (DPO).

• Quantized and compiled to ONNX. Served at the edge to minimize inference cost.
• Pipeline written in Rust with Python bindings (PyO3). Additional Rust-WASM in browser.