AI Development Intern

Overview

In Summer 2025, I joined Hovian Inc., a startup dedicated to building an AI-driven calendar. Working in a tight-knit 5-person team, my focus was building the AI Image Generator module.

The system allows users to generate GIFs on demand. I built the Python backend that orchestrates generation, allowing users to toggle between Midjourney's Flux model and OpenAI's GPT-4o depending on their specific needs.

Challenge: The Cost of Fidelity

As a startup, budget constraints were tight. Generating high-resolution images for every single frame of a GIF was prohibitively expensive. We needed a way to produce visually appealing results without burning through our compute budget.

I proposed a creative solution: use lower-precision base models and apply a pixelation filter to mask the imperfections. This created a cohesive "retro" aesthetic that became a signature style feature, rather than just a cost-cutting measure.

Porting Libraries with Gemini CLI

To implement this pixelation effect, I found an open-source library called pixelit. However, it was written entirely in JavaScript for frontend use, which was incompatible with our Python backend.

Leveraging the Gemini CLI (in early 2025, before the "AI Agent" concept had fully matured), I reverse-engineered the JavaScript logic and reimplemented the core algorithm in Python. This allowed us to perform the image processing efficiently on the server side.

Balancing Act

The most demanding aspect was the timeline. The project development coincided perfectly with my university finals week. I had to rigorously manage my schedule to revise for exams while developing this feature.

Outcome: I successfully delivered the backend by the late-June deadline, enabling the product launch, and simultaneously achieved a 4.0 GPA for the semester.