Intelligence Admin + Iceberg.

Shipped a Claude Opus 4.7 + pgvector + Voyage embeddings admin surface where every generated brief, article draft, and communication draft passes through a voice-compliance wrapper that pre-injects the system prompt, post-scans output for Iron Rule violations, and retries up to twice with explicit correction instructions.

The hardest call was the Iceberg Fallout Mapper: it takes a decision, asks what falls apart, maps ten zones of fallout, stress-tests against an adversarial reader, surfaces sanitizing language in planned communications, and generates authentic replacements. Shipping the adversarial reader meant building a prompt that could pressure-test the author rather than flatter them.

I would start with a harder evaluation harness. Voice-rule enforcement at the API layer works, but I lack a regression suite that catches drift in the rubric itself; the rubric evolves, but I have no ground-truth dataset to confirm a rubric update did not degrade prior outputs.