SermonAI

Pastors spend hours every week manually repurposing one sermon into summaries, discussion guides, devotionals, blog posts, and verse lists — then spend more time editing AI output that loses their theological voice. The product had to handle messy real-world inputs (raw audio, shaky video, pasted manuscripts), generate outputs that felt like the pastor rather than generic AI slop, and operate reliably enough for a weekly ministry rhythm.

Started with the intake problem: pastors don't prepare the same way, so the product had to meet three distinct starting points — audio/video upload, YouTube link, and manuscript. Mapped each path through transcription → structured generation → human review → publish, treating review as a required UX step so AI never publishes without sign-off. Kept transcription and generation as separate, retryable API routes so failures are contained and the UI stays honest about progress. Applied tone-preservation as a product constraint from day one, not a post-hoc filter.

Next.js App Router monorepo (Turborepo), Supabase-backed persistence with typed client generation, Deepgram for transcription with retry handling, OpenAI structured outputs for five content types, Stripe billing with trial and subscription paths, team workspaces with role-gated access, and next-intl for localized strings. Marketing site tells the same story the in-app product delivers.

Live production SaaS with active ministry team subscribers. Subscription and trial path in place for recurring use. One upload generates five ministry-ready outputs per sermon. Full design and development owned end-to-end — problem framing, user flows, API architecture, billing, and ongoing iteration.