Philip Roy, applied AI in production. Ask the pool.

Multi-agent content production

Forward-deployed for paying clients and running around the clock. A dispatcher hands work to voice-bound drafters; a blind cold-read agent and a five-dimension quality gate review every draft; bounded auto-revision repairs against named strategies, capped before a human steps in. High-volume voice-matched production across many executive voices, every post dispositioned before the next batch goes out. The core production prompt has been rewritten through 22 documented versions, each one a caught failure fixed.

Persistent memory + personality

A verbatim knowledge graph, behavioral-drift tracking, and an experiential layer that writes a short poetic encoding of how each session felt to work through, so an agent carries both the facts and the texture of the work across hundreds of sessions instead of starting cold. Retrieval is ranked by a seven-signal reranker, and a measured personality layer keeps the agent in character far more consistently than the base model. 34,782 entries live in production.

Deep Applied Research

A protocol that runs research agents through structured hypothesis-to-verdict cycles, holding each finding to a baseline before it counts as a result. It also audits itself: it caught its own verdicts skewing optimistic, adopting 77% of the reports it reviewed, so it named the bias, corrected it, and re-measured.