Work

A 6-domain production platform spanning CoreOps (cyber-physical RE), DevOps (FORGE software factory), and 4 additional specialized domains. 18 autonomous agents run on coordinated cron schedules: nightly research scans at 0200, automated builds from 0300–0600, daily AI-voiced podcast at 0645, and continuous health monitoring. The entire system was built through hundreds of iterative sessions with Claude Code.

Key patterns: agent-per-script (no frameworks), tool-use over prompt chains, circuit breakers with 12-crash threshold, metrics integrity watchdog, deterministic provenance signing.

Answers the question most organizations skip: which tasks should humans do, which should agents do, and which need human-in-the-loop? A 7-step workflow from meta-model scanning through role discovery, task decomposition, and human/AI classification to actual MCP server code generation. Includes a 12-question Task Delegation Wizard and multi-device facilitation mode for live workshops.

Integrates NIST AI RMF, EU AI Act, ISO 42001, MIL-STD, and DO-178C. Model Authority Engine cross-references frameworks automatically. RACI matrices with AI as first-class participant.

Semantic classifier via nomic-embed-text embeddings routes prompts across 5 local Ollama models based on task type, complexity, and GPU load. Per-task-type thresholds (draft 0.50, review 0.75, architect 0.80) ensure quality isn't sacrificed for cost. 62K tokens saved in production with streaming output and batch mode support.