PointCard (Y Combinator W19)

• Designed and built the company’s first MLOps platform in AWS, enabling scalable, end-to-end workflows across both staging and production environments.
  
  ⭐️ Created a real-time inference engine using FastAPI, Docker, and Kubernetes, alongside batch pipelines
  orchestrated with Apache Airflow and custom operators
  ⭐️ Used SageMaker for model development, with MLflow for versioning and registry management
  ⭐️ Deployed containerized apps via ECS and ECR, and implemented CI/CD pipelines with GitHub Actions
  ⭐️ Built a private Python package repository with CodeArtifact, and provisioned cloud infrastructure using Pulumi (IaC)
  ‍
• Configured a Databricks Spark cluster for large-scale ETL pipelines processing terabytes of data from Redshift and S3 (Avro, Parquet, JSON), reducing manual overhead by 11%.
• Integrated Kafka consumers and Spark Streaming to support real-time data ingestion and processing.

Block, inc. (fka Square, inc.)

• Designed and built a deep learning-based recommendation system in PyTorch to serve millions of sellers with personalized product suggestions. The two-tower model architecture (user + product) featured:
  ‍
   ⭐️ A user tower that processed engineered features and sequences of past actions fed to an RNN.
   ⭐️ A product tower that used a fine-tuned BERT encoder on textual descriptions.
  ‍
  The model was trained on NVIDIA A100 GPUs using a custom information retrieval loss function implemented from scratch based on academic research. It achieved a 15% lift in A/B testing over existing ranking methods, contributing to $5M+ in additional annual revenue.
  The system became the basis of a patent and was presented at NVIDIA GTC 2021.
  ‍
• Created a Python package designed to vectorize any text or images, using Transformer-based models for NLP and Computer Vision for Tensorflow/Keras and PyTorch. It was containerized, deployed as a microservice and made available via a REST API.

• Led the development of an internal deployment platform that streamlined ML serving by automatically converting trained models into FastAPI-based servers, containerizing them with Docker, and deploying them to Kubernetes clusters on GCP, significantly reducing engineering overhead and time-to-production.
• Developed a LightGBM-based cross-sell model that prioritized high-propensity users, driving an 18% increase in sales conversion rates. The model surfaced daily ranked opportunities to account managers via Airflow, enabling more targeted and effective outreach based on predicted conversion likelihood.
• Authored a novel deep learning-based survival analysis model, using PyTorch, designed to compute when users would repay their Square Capital loans in full.
• Prototyped a novel deep learning-based model for customer support that ranked the most relevant solutions to a user's inquiry by leveraging word embeddings from the submitted email questions.