Luca – Python, Machine learning, PyTorch

It's an end-to-end ML platform that covers serving, training, and feature store infrastructure.

• Shipped the cross-team rollout of the company's unified Kotlin serving platform.
• Spearheaded the adoption of Triton as the runtime for serving production BERT models in-house.
• Led efforts to keep Ads model latency under SLA while the neural network scaled in size.
• Retrained the pClick model without JSON feature parsing and introduced server-side batching to reduce latency.
• Scaled the Feature Store built with Go and Redis to support NN-era throughput demands.
• Identified and led the replacement of OpenAI with task-specific BERT models for content moderation.
• Designed and built DNN training infrastructure for billion-row datasets using 8-GPU data parallelism.
• Rewrote the GNN embedding pipeline end-to-end in PyTorch-Lightning based on TwHIN and TransE papers.
• Prototyped Spark offline joins for pre-trained embeddings at scale.
• Sponsored the development of the Multihead DNN architecture to consolidate Feed model pipelines.
• Migrated XGBoost pipelines to distributed training on Spark.
• Defined the multi-year training platform strategy, including GPU serving, orchestration evaluation, and Kubernetes migration.
• Grew the CoreML team through hiring and promotions while maintaining near-zero attrition.

It's an ad platform for the company that makes and sells smartphones, computers, tablets, wearable devices, and digital services like music, TV, and an app store.

• Acted as DRI for about 10 engineers migrating preprocessing jobs to Airflow.
• Kickstarted the internal MLOps team.

It's a model platform and online feature store that provides training and serving infrastructure, ensuring ML training-serving consistency.

• Implemented core components of the company's Online Feature Store and the Model Platform.
• Led the Model Platform initiative in the Core-ML team.
• Owned infrastructure to train and serve ML models in Spark, XGBoost, and Scikit-learn using MLeap as the serialization format.
• Led a cross-team project to implement the Online Data Store supporting ML online predictions for Yelp ML teams.
• Trained a Random Forest model for Store Visits predictions and improved accuracy over heuristics.
• Built real-time streaming ML pipelines in Apache Flink for geo-clustering and ML predictions.
• Coordinated open source contributions to spark-redshift-community, MLeap, MLFlow, xgboost, and xgboost-predictor.