Ioan Popa / Principal Product Designer & AI LeadCisco · Splunk / California

Designing AI,
from the inside out.

I design interfaces, build working prototypes and fine-tune models for hard problems that matter.

Selected work.

Making something, finding an answer, restoring a memory. Three different problems that need room for human judgment.

Build workspace: reconstructed design wireframe. Open the case study for details.

Self-initiated · Design & build

Cozonac

A visual AI builder designed around the second, third and twentieth decision—not just the first generation.

Product design & implementation

Working editing and recovery, demonstrated in the case study recording.

Inside the project
Memory Bridge: reconstructed design wireframe. Open the case study for details.

Self-initiated · Working prototype

Memory Bridge

Restoring family photographs means designing for the moment someone says: that doesn’t feel like them.

Product design & prototyping

Actual project images and a working review-before-save flow.

Inside the project

Earlier work.

The foundations of this work: enterprise platforms, mobile products and the problems that shaped my practice.

Design practice.

Across these projects, I map the difficult moments and build prototypes to test them: keeping context, correcting a result and knowing when to ask for approval.

Model training.

Sometimes the problem sits inside the model. I work on anatomy and cinematic feel, and fine-tune the text encoders to interpret lenses, lighting and scene direction.

Explore the image studio
12BModel_mini
Working preview

Text to image, with references.

I added reference support and fine-tuned its 6B language encoder. This is currently my strongest model for image quality.

07
Working image references9-reference binding in developmentLoRA adapters merged into the model
Lens + lightScene direction6BFine-tuned text encoder12BImage model
32BWork in progress

Fine-tuning the visual language.

I fine-tuned the image model and its Mistral text encoder, and integrated reference support.

11
Image referencesLoRA adapters → merged into the model
Lens + lightScene direction24B MistralFine-tuned text encoder32BImage model

Image Studio.

I built a Diffusers studio that brings prompts, @image references and @lora direction into one place.

Image Studio
Canvas
Original test prompt12B image model
6B fine-tuned text encoder

Cinematic close-up of a woman in her late forties sitting beside a rain-covered train window at dusk. Her face is turned slightly toward the glass, eyes looking outside. Cool evening light shapes her face; a warm reading lamp provides a subtle edge light. Shot on an 85mm lens at f/2, focus on the near eye, naturally falling depth of field. Visible pores, fine facial lines, restrained colour grading, candid expression.

Actual model output · recorded example

The same four prompts, run on Model_mini and Model_max. Select a scene, then switch models to compare saved outputs. No live generation runs here.

About Ioan.

Principal Product Designer & AI LeadCisco · Splunk / California

I’m drawn to hard problems where design can make a difference.

I grew up in Romania with two teachers for parents. Explaining complicated things clearly still feels like the family business.

There’s another side to this portfolio. Same curiosity, considerably longer exposures.

A conversation is a good start.

Something worth
working on?

1ONICA FM

Choose a perspective. Access is approved personally.

Public mode is open to everyone. Other perspectives open by invitation.

Level 010 XP

100 XP to the next level

Your progress stays in this browser.