Founder voice system
A voice-fidelity content engine for a teaching brand
A dual-pool RAG system that turns a teacher's 130-page book, 15 articles, and 100 episode transcripts into voice-consistent drafts across three audience tiers — without losing the founder's argument or register.
Mock metrics
- 10hrs
- of writing time reclaimed weekly
- 130
- pages of source voice indexed
- 3
- audience avatars calibrated
The challenge
Generic AI stripped the founder's voice. Scaling content meant losing the philosophical precision that made the brand worth following.
Three audience tiers — beginner, intermediate, advanced — each needed a different register. One model holding all three produced mush. Manual writing across a sprawling corpus took hours, sometimes days, per piece.
The build
Dual-pool RAG on Supabase + pgvector: one pool for voice and framing (book, articles, 100 indexed transcripts), one for citations. The pools never mix, so the model sounds like the founder and cites the right sources.
n8n orchestrates a clarifying-question loop, avatar routing across three tone profiles, Claude Sonnet 4.5 drafting through OpenRouter, and a deterministic Notion review chain. A custom markdown-to-Notion conversion layer handles 10,000+ character drafts in chunks. Hash-deduped quote ingestion keeps the vector store clean on re-runs.
The outcome
What used to take hours of cross-referencing the corpus now takes minutes of conversation. The founder spends time on judgment and argument, not blank-page assembly.
Drafts arrive in the founder's voice on the first pass, calibrated to the right audience tier. Review becomes editing instead of rewriting.
Build the real story
Have a workflow worth turning into an asset?
Share the shape of the system you want to build. The full contact page will collect the deeper context.