Build for editing
Generated material remains structured so people can keep working instead of accepting or rejecting a finished image.
Epsilon / AI-assisted creative tools
More output. Still editable.
I lead technical architecture and built much of the interaction and front-end foundation for AI-assisted creative tools that keep generated work editable, understandable and ready for review.
- Role
- Staff Software Engineer / Tech lead
- Contribution
- AI interactions, visual editing, state architecture and implementation
- Collaboration
- Product, creative, design, ML and engineering
Interaction
AI-assisted workflows with visible state and human decisions
System
Editable visual work across static and motion formats
Quality
Review and evaluation states that make changes easier to understand
The problem
Creative work needed more leverage without less control.
AI could help people explore and adapt visual work, but generation alone was not enough. The product also needed to preserve editing, reveal important changes and give people a dependable way to review the result.
I kept coming back to one question: “How can AI help while keeping people close to the decisions?”
The product
One understandable path from direction to review.
I helped shape this as one connected product. Each step keeps the material editable and makes the next decision clear.
- Set the direction People provide goals, constraints and approved material.
- Explore options AI assistance creates possibilities people can compare and guide.
- Keep editing Chosen work remains structured instead of becoming a dead-end result.
- Adapt and review Related formats stay connected while people inspect what changed.
People stay in the loop throughout the AI-assisted work. Evaluation states surface drift and failure before release.
Key decisions
Make the complicated parts easier to trust.
Synchronize the formats
A shared state model carries changes across sizes and motion output while preserving the ability to adjust individual formats when needed.
Make AI state visible
Clear states help people understand what the system is doing, where it needs attention and what can happen next.
Treat evaluation as product work
Layout, color and visual comparison checks feed visible review states. The goal is not just to generate—it is to help people know when the output is ready.
My contribution
I helped turn the idea into software the team could extend.
I worked across product definition, interaction design and production engineering. I lead the technical architecture, implemented much of the core front-end experience and helped other engineers carry the work forward.
Shaped
Various AI interaction models, including guided changes, synchronized formats and clear review states.
Built
Much of the canvas, state architecture, AI-assisted interactions, review system and evaluation loop.
Connected
Product, creative, design, ML and engineering partners around a shared model of how the platform should work.
Extended
The work beyond one launch by mentoring engineers and turning repeated needs into reusable patterns the team could build on.
What changed
People could generate, edit and review the work in one flow.
The product moved beyond generation: people could change the result, understand what happened and review the work before it moved forward.
Generation is only the beginning. A useful AI product gives people editable output, visible state and a dependable way to review what happened.