PRISM
An AI content engine that turns a marketing strategy into research-grounded social and email content, scored against each brand's voice and mission, and held for human approval before anything publishes.
Application
Brands — the pipeline's home base: every active brand, its channels, and how much Voice/Mission config backs it
Scheduling and Publishing
Scheduling — approved content gets an exact publish date and time, set by a human
Toolkit
I built PRISM to make AI content I could actually trust at volume. It routes each stage to the right model, scores every draft against a real rubric, and never publishes without my approval.
The Context
PRISM runs on a Brand → Voice → Mission hierarchy, configured per brand: identity, persona, and audience live in the database. Every piece starts with live research on its specific topic, shaped by a structural template built for its specific channel.
The Approach
A model router assigns each pipeline stage to the model built for it — Perplexity for live research, Anthropic for on-brand drafting, Google Gemini for the media plan, Ollama for zero-cost scoring — across Ideation, Research, Generation, Scoring, Review, Approval, and Publish. Every piece is scored against a real rubric before a human decides.
What I Built
An end-to-end pipeline powering real content production across three active brands — including Right Question and Lenormand Ai — with 31 per-platform templates spanning six channels, a Supabase-backed Brand/Voice/Mission system, and live publishing hand-offs to Postiz and MailJet. Cost is tracked per piece.
The Outcome
A real, in-production internal system — every piece researched, scored, and reviewed before I approve and schedule it, all the way through to live posts across real brands. The full pipeline already runs end-to-end; current work is tightening research quality and closing the loop with post-publish analytics feedback.