Duarte AI

I joined Duarte to help build and maintain Duarte AI, an AI-powered web application that assists users in creating persuasive presentations.

The application was built as a Next.js app hosted on Vercel, with two supporting microservices hosted on AWS ECS.

One microservice, the exporter (written in Python), converted LLM-generated slides into PPTX presentations. When I joined, this service was broken and only partially implemented. I refactored and completed it, which required in-depth research into the structure of PPTX files and how to programmatically create and manipulate them.

The other microservice, the converter, transformed PPTX and PDF files into images. It used LibreOffice to convert PPTX to PDF and ImageMagick to render PDFs as slide images. We encountered an issue where fonts in the slide images did not match the originals, causing text overlap. After extensive research, I implemented two solutions: updating LibreOffice to a beta version to resolve an embedded font issue, and creating a font substitution map to replace common fonts with equivalent free alternatives.

My most impactful contribution was refactoring the application to enforce LLM output adherence to JSON schemas. Previously, the LLM produced Markdown, and the application attempted to parse the required information from it, which was error-prone and unreliable.

JazzHR: Text2Apply

At JazzHR (Employ Inc.), I collaborated with a team to develop a new feature called Text2Apply. JazzHR is an applicant tracking system that provides tools for companies seeking to hire new employees. Text2Apply enabled candidates to apply for jobs via SMS. My primary responsibility was implementing a REST API client to interface with the SMS service, an internal service maintained by another team at Employ. I followed the documentation and regularly consulted with the SMS service team to complete this task. Additionally, I was tasked with ensuring 10DLC compliance and contributed to the development of related front-end React components. I worked with my team to design the overall API and database schema for the feature, as well as to build automated tests and a test database.

Platform.CC & PBrew

While at Contextual Code, we migrated all our web applications to Platform.sh, a managed hosting provider that automatically builds the hosting environment from a configuration file. For developers at Contextual Code, spinning up a local development environment was always a challenge, and there was always uncertainty whether changes that worked locally would function correctly in production. I realized that the Platform.sh configuration file provided detailed information about the hosting environment, making it possible to build a tool that could recreate that environment locally using Docker. Over the course of a weekend, I built a working prototype. It was well received by my colleagues, and I was given the greenlight to continue development. It eventually became a crucial tool for all developers at Contextual Code. It was named Platform.cc, with CC standing for Contextual Code, a play on the service it was based on.

Around this time, many developers at Contextual Code switched to M1 Macs. Docker had poor performance on these machines, so I decided to build a new tool that could operate without containerization. This tool, named PBrew, used Homebrew to manage dependencies and made it possible to install multiple versions of all the software needed for a local development environment. It mapped all services to custom ports to ensure multiple versions of the same software could run simultaneously. While this approach sacrificed parity with the Platform.sh production environment, it was very fast and made local development significantly more efficient.

https://github.com/chompy/platform_cc https://github.com/chompy/pbrew

OncoLife Survivorship Care Plan

A web application that helps cancer patients identify optimal treatment options. I collaborated with a fellow developer and a team from the University of Pennsylvania to build a questionnaire driven by a sophisticated logic engine. This project involved rebuilding their existing ColdFusion system using PHP with the Symfony framework and the Ibexa CMS. We leveraged the CMS to define the nodes within the logic engine, which typically represented questions and answers for patients or their care providers. Each node included conditional logic to dynamically determine which subsequent questions and answers to display. Patient responses were then used to generate a comprehensive report, also constructed using the same logic tree system. Later, we expanded the application to interface with FHIR APIs and developed a containerized version for distribution to third-party partners. Through rigorous automated testing, we successfully launched the application with no significant bugs or issues.

https://oncolife.oncolink.org

Modern Language Association

This was my first major project at Contextual Code. I collaborated closely with both my colleagues at Contextual Code and directly with the Modern Language Association (MLA) to build and maintain their website.

The site was powered by the Ibexa CMS (formerly eZPublish) and featured a sophisticated member signup flow, plus a full-fledged bookstore that blended Shopify integrations with custom e-commerce solutions. Additionally, we worked with the MLA’s own developers who built a custom API that allowed us to fetch and display data from the MLA’s internal database.

https://www.mla.org