01

Context

AdaIA began during an internship at Sparkling Tech with a question:

Can a conversational interface detect how someone is feeling and respond with structured empathy? The resulting system is an emotional well-being agent based on OpenAI Assistants v2, fine-tuned on GPT-3.5, with a React-based chat interface and the ability to export conversations to PDF. It’s deployed. It works. I built it.

02

The challenge

Designing the experience of a therapeutic conversation is different from designing any other interface.

The user isn’t buying something or browsing a catalog—they’re expressing how they feel. Every design decision had to support that moment: the typography, the response pace, the typing indicator, the silence before the agent replies. And at the same time, the system that generated those responses had to be trained, deployed, and maintained by the same person who designed the interface.

03

My process

I started with the emotional model.

Before designing a single screen, I researched Plutchik’s and Ekman’s models to understand what the agent needed to detect—not what emotions existed, but which ones it could recognize in the text and respond to appropriately. That research became the training datasets: conversation pairs in JSONL format where each of the agent’s responses had structure, empathy, and emotional accuracy.

The fine-tuning was performed in the OpenAI Dashboard with multiple validation iterations. Each version of the model was tested in the Playground by comparing A/B responses—which one detected emotion better, which one responded more consistently, and which one avoided empty generalizations.

The interface design prioritized what Lupton would call the experience over time: not just how the screen looks, but how it feels to use it second by second. The “typing…” indicator isn’t just for show—it captures the rhythm of a real conversation. Exporting to PDF isn’t just a technical feature—it’s a way to save a record of what happened.

04

Results and Lessons Learned

AdaIA is in production.

The chat works, the agent responds with emotional structure, conversations are exported to PDF, and the system has a login feature. The remaining tasks—multi-user support and audio input via Whisper—are documented and ready for the next iteration.

This project demonstrated something that wasn't in the brief: that designing the user experience for an AI system and building that system are one and the same task. There's no handoff between the designer and the developer when you're the same person.

• Meet AdaIA

01

A conversational interface is designed in time, not in space. Pacing is just as important as layout.

02

Training an AI model is a design process—each data pair represents a decision about what the system should say and how it should say it.

03

Documenting technical roadblocks as a reusable repository turns errors into infrastructure.