fka Square: Disputes & Risk Infrastructure
• Worked one-on-one for an hour+ with more than 10 technical and non-technical peers to teach effective use of coding agent tools across domains including frontend and backend software engineering, personal tool development, production design, brand design, and animation.
• Productionized inference pipeline to evaluate if additional information was needed from a seller according to policy to win a dispute, achieving 90% precision on evals, improving upon an ops-driven review process by reducing cost and increasing throughput.
• Hosted 50 weekly, hour+ hack sessions pushing the limits of agent coding tools, building software live using spec-driven development and long-running agent loops that ran autonomously until validation criteria passed.
• Co-hosted weekly company-wide AI working group, socializing best practices and giving engineers a platform to present their work with AI and language model tools to a broader audience, regularly attended by 50+ engineers
• Established first-in-kind cross-business-unit Kafka integration in Ruby, enabling more than 1,000 microservices to consume previously inaccessible streamed data.
• Built and deployed internal router and dashboard enabling low-friction prompting, experimentation, and prototyping across different LLMs and providers, lowering the barrier to learning and empowering the experimentation that led to the disputes document intelligence work.
• Built and delivered menu text extraction tool for Spanish-language menus, creating a first-pass automation for a previously manual process.
• Piloted structured data extraction from dispute evidence documents submitted by sellers, achieving 90% accuracy across 11 fields with LLMs, used to aid in constructing processor responses for challenged transactions.
Technologies:
• Java
• Guice
• S3
• Kafka
• DynamoDB
• Ruby
• Rails
• Python
• FastAPI
• Datadog
• Terraform
• AWS Bedrock
• Databricks
