Research Output
Publications
Whitepapers, technical briefs, and field-tested architectural methods.
Claude + Docling: Local Document Intelligence at Scale
How we indexed 400+ personal documents — financial, medical, vehicle, and household — into a persistent, privacy-preserving local cache using Docling MCP and Filesystem MCP. Covers the 3-phase optimization journey from naive batch processing (zero success) through expensive individual conversion to the hybrid pre-scan strategy that achieved 95% token efficiency and 99.2% reliability.
Key Findings
MCP Composability Gap: No standardized file passing between MCP servers operating in different contexts (user computer vs. Claude computer). Docling MCP cannot consume files from Filesystem MCP paths — forces architectural workarounds.
macOS .DS_Store: Hidden Finder metadata files break batch conversion 100% of the time. Pre-scan detection + auto-fallback to individual mode is mandatory.
Cache Persistence: Docling's local cache survives Claude Desktop restarts and conversation boundaries — enabling true cross-session document memory.
Privacy Architecture: Zero cloud upload. All processing local. HIPAA, GDPR, and attorney-client privilege compliant by design.
Cross-MCP File Processing: Lessons from sort-O-matic
Honest documentation of what didn't work: building content-aware document organization across the Filesystem MCP / Docling MCP boundary. Discovered that cross-computer file passing has no standardized solution. Pivoted from content-based to filename-based classification — shipped the 80% that works, documented the 20% that's blocked.
Recommendations Filed
• Docling MCP: Accept base64 file content as alternative to file paths
• Claude.ai: Unified file handle passing between MCP servers
• MCP Protocol: Standardized cross-server file transfer pattern
Performance Brief — February 2025
O-Matic Skills Performance Report
Quantified comparison of convert-O-matic + sort-O-matic vs. baseline Claude, Filesystem-only, and Docling-only approaches. Includes the Pets folder case study (14 files organized in 2 minutes, 3,700 tokens) and Medical folder projection (300-500 files, 10-15 minutes, HIPAA compliant).
Token Economics
- Batch (clean dir): ~150 tokens
- Individual (100 files): ~12,000 tokens
- Pre-scan hybrid: ~300-750 tokens
- Manual upload: ~50,000+ tokens
MSP ROI Projection
- 50 clients × 100 docs × 4x/year
- Without skills: 700 hours ($105K)
- With skills: 23 hours ($3.5K)
- Annual savings: $101,500 (97%)
Featured Whitepaper
Closed Factory 2.0 — Multi-Agent Blueprint
Updates the Closed Factory model for next-gen agents, introducing the Spec X governance layer, enhanced multi-agent orchestration, and the Brandy / Probot / Carver roster.
Foundational Whitepaper
Closed Factory & Multi-Agent Blueprint (Original)
The original Closed Factory whitepaper. Core governance model, deterministic agent handoffs, Spec-10 persona architecture, and the principles behind safe multi-agent collaboration.
Upcoming Publications
MCP Composability: Cross-Server File Intelligence
Proposed patterns for unified file passing between MCP servers. Recommendations for Docling, Filesystem, and protocol-level improvements.
Spec-10 Architecture Overview
Structure, stability, identity, and cross-agent integrity.
Dynamic Intent Mesh
Intent routing, context binding, drift reduction, and evaluation layers.
Human-AI Roundtable Methods
Debate frameworks, cross-validation protocols, and multi-lens reasoning.
Agent Persona Stabilization
Reproducible identity systems for long-running agents.
Applied Governance Models for AI Systems
Operational practices for safety and auditability.