FlowState Intelligence.

Shipped a Next.js 16 + Supabase (pgvector, RLS) + Claude Opus 4.7 platform that scores organizational friction across four types, writes 30+ intelligence tables, and surfaces a three-surface admin where an operator can browse, chat, and generate content grounded in live captured data.

The hardest call was the reflexive rubric scoring pattern: Claude grades its own prior output against an explicit rubric with a retry loop and a defensive em-dash scrub, so voice rules apply at the API layer rather than as post-hoc cleanup. That pattern now runs in three contexts (book QA, question validity, communication drafts) and is being extracted into a reusable skill.

I would change the intelligence schema up front: early tables optimized for weekly human synthesis and the pgvector path was bolted on later. Starting with semantic indexing as a first-class primitive would have saved two weeks of migration.