Creatiz.ai

Creators were stitching together four or five separate AI tools to produce a single piece of content — one for ideas, another for scripts, another for captions, another for repurposing. Every handoff meant copying context between tools that knew nothing about each other.

The opportunity was an AI content OS: one workspace where the model already has your voice, your audience, and your previous work as context, so each step builds on the last instead of starting cold.

I built the original platform from scratch — the content generation pipeline, the workspace interface, and the LLM orchestration behind it. It worked and it validated the concept, but it was a first version in the way first versions always are: functional, rough at the edges, carrying decisions made under time pressure.

That version did its job. It proved people wanted the workflow, which is the only question a first version needs to answer.

I later rejoined the team as a full-stack developer on the platform I had originally written — an unusual position that meant I knew exactly which early decisions needed to be undone and why they had been made.

The frontend was rebuilt on a proper component system, cutting page load time by roughly 40%. New AI features shipped on top of a cleaner generation pipeline, with model routing so simpler operations stopped paying for the most expensive model available.

Most AI product work is not prompt engineering. It is streaming responses that feel responsive, handling model failures without dead-ending the user, keeping inference costs proportional to revenue, and building an interface that makes a non-deterministic system feel dependable.

Having built the first version and then lived with maintaining it is the most useful experience I bring to other people's AI products — I have already paid for the shortcuts.