Period: September 2025–May 2026
Building three applications for generating, managing, and experimenting with campus-specific AI humor.
Crackd.ai generates image-based captions for the communities at Columbia, Barnard, and Stanford. Its humor styles, called “flavors,” combine images with campus terminology, current events, and voices inspired by fictional characters and public figures.
The wider platform draws on regularly updated data, including campus activity and Sidechat discussions. That context helps it generate references its audience can recognize.
My involvement developed in two phases: contributing to the shared platform within a larger team, then independently implementing three deployed applications against an existing live database.
September–December 2025
I worked across the UI design, admin, and prompt-engineering pods.
I implemented interface changes based on the design team’s work, developed admin features for inspecting platform activity and statistics, and experimented with new humor flavors through the chaining matrix.
We organized work in Linear and collaborated through GitHub branches and pull requests. I coordinated changes with other contributors, reviewed feedback, and resolved merge conflicts within an actively developed codebase.
January–May 2026
In the following semester, I built my own implementations of the user application, admin application, and LLM chaining matrix while a separate team continued developing the main platform.
I implemented each application end to end, including interfaces, authentication and access controls, application logic, integration with the existing database and services, and deployment. These are functioning applications connected to live data.
An application for uploading images, generating captions, browsing existing submissions, and voting on image-caption pairs.
I built the user journeys around generation and feedback, including upload handling, progress states, caption viewing, voting rounds, personal caption collections, and highlights.
An interface for inspecting and managing the records and configuration behind the platform.
I built searchable, filterable views for captions, images, profiles, and humor configuration, alongside permissions management and analytical views. Administrators could examine individual records, inspect prompt-step details, and explore patterns in caption feedback.
A development environment for building and testing multi-step humor-generation workflows.
I implemented flavor browsing, ordered prompt-chain configuration, model and input/output settings, test-set selection, run progress, and caption review. Outputs from earlier steps could be referenced by later prompts, allowing generation to be broken into distinct operations.
A flavor is more than a single instruction to “be funny.” A chain can separate image interpretation, contextual information, and caption writing, with intermediate outputs informing later steps.
Building the matrix meant making those workflows configurable and inspectable. Using it for prompt experiments meant considering which examples, instructions, and processing stages helped produce a particular voice or style.
Working across all three applications changed how I thought about an AI product. The generation matters, but so do the interfaces for using it, the tools for maintaining its data, and the environment for investigating why an output changed.
Humor adds another complication: technically correct is a fairly low bar for a joke.
Coursework with Professor Lydia Chilton and Michael Roger.
