I design AI systems that get out of the user's way.

Separated transcription from generation in SermonAI so the UI could be honest about pipeline progress.

When the AI is doing multiple sequential jobs, combining them into one opaque step erodes trust. Keeping them distinct let the interface show real state — not a spinner lying about how far along we were. Designing a non-deterministic, multi-stage AI workflow means the UI must expose what the system is actually doing.

Designed every SermonAI output to land in a reviewable state first — the AI never publishes without human sign-off.

Automation complacency is a documented HCI failure mode. When users trust AI output without reviewing it, errors compound silently. Building in a mandatory review step isn't friction — it's the design preventing a category of harm specific to tools that generate content in someone else's voice.

Chose no-code tooling for Swyvvl to validate the product before committing to a custom stack.

Webflow + Xano + Wized was the right call for a pre-revenue product with unproven assumptions. The lesson wasn't that no-code fails — it's knowing exactly where the ceiling is before you hit it, and scoping accordingly.

Built this portfolio in Next.js + Supabase rather than Webflow to demonstrate the same stack I sell.

If I claim I can ship production Next.js AI products, the credibility check is whether I'm willing to build my own tools with that stack. This site — including the retrieval (RAG) pipeline behind the upcoming work-page personalization — is the answer.