• Led a team of five undergraduate students to design and deliver the PainUSA non-profit website, coordinating milestones, delegating technical tasks, and aligning design and client requirements.
• Guided user interface prototyping in Figma and implementation of the public website in Webflow.
• Engineered and deployed a Mapbox-based interactive clinician lookup map, leveraging Cloudflare R2 for data storage and hosting, and embedded HubSpot forms for newsletter sign-up.
• Managed client communication with Stanford Division of Pain Medicine leadership, producing structured handoff documentation to ensure sustainable post-delivery ownership.

• Upgraded Ternary’s cloud cost forecasting API by integrating Meta Prophet, enabling automated daily and weekly client spend predictions and improving modeling sophistication beyond the prior linear regression baseline.
• Improved forecast reliability through time-series cross-validation, hyperparameter tuning, and confidence interval integration.
• Engineered backend workflows connecting Go and Python services, implementing schema-validation tests and CI pipelines to ensure forecasting accuracy and production stability.
• Deployed and maintained staging environments on Google Cloud, using Terraform to provision and manage shared resources and support pre-production UI testing.
• Enhanced the frontend user interface with a rotating slideshow of interactive line charts, visualizing top cost-contributing business dimensions.

• Automated textual and layout segmentation of historical Detroit Free Press front pages, extracting article-level structure such as titles, bylines, body text, and reading order from JP2 images and OCR data.
• Developed an MLOps pipeline integrating large language models, Gaussian mixture models, and computer vision for text-type classification, article boundary detection, and noise labeling.
• Validated segmentation and classification performance through controlled A/B testing, comparing pipeline variants to optimize accuracy, robustness, and processing efficiency.
• Reduced segmentation cost by 50% compared to manual labeling while increasing throughput to 12 pages per minute.

If the embed doesn’t display on some browsers, the file can still be downloaded: Open PDF in a new tab.

• Enhanced the retrieval augmented generation pipeline powering an in-car voice assistant for Toyota vehicles.
• Developed a new warning-light query handling module and deployed multiple system iterations through vehicle simulator testing.
• Implemented multi-intent recognition, refusal logic, and query translation components to improve accuracy, safety, and user interaction quality.