At RepairPal, I built AI diagnostic recommendations for the marketplace using OpenAI, structured outputs, and retrieval over repair data. I added caching, output validation, and fallbacks to keep costs predictable and prevent bad responses from reaching customers.
I built the feature from the React and Next.js front end through the Node.js APIs, lifting conversion on the busiest marketplace pages by roughly 8%. I also set up quality checks and an evaluation suite to compare model outputs with known-good cases.
At Sanebox, I prototyped LLM-based email categorization and summaries alongside the existing ML pipeline, and improved the categorization pipeline to cut latency by roughly 40% for high-volume mailbox processing. I also reworked the first-run setup flow, lifting new-user activation by around 15%.
At Cadre5, I built and maintained full-stack features for enterprise and government clients, including REST APIs with role-based access control and screens brought to Section 508 accessibility requirements. I also helped migrate older pages to a React single-page front end.
