Selected Work

The system orchestrates multiple specialist agents working in a daily pipeline. Each agent focuses on a specific behavioral signal:

Signal Agents: - S3 (Ritual Language): Pattern-matching agent detects lexicon use and cultural markers - S4 (Meaning Making): LLM-powered depth analysis evaluates comment substance and contribution quality - S5 (Authenticity): Spam and bot detection agent filters synthetic engagement - S6 (Social Gravity): Connection analysis measures influence and network position - S7 (Participation Velocity): Momentum tracking identifies rising contributors

Weekly Synthesis: A separate identity agent generates natural language profiles for high-value participants, summarizing their engagement patterns and community role.

Coordination: All agents write to a central PostgreSQL database. A VPS-based cron orchestrator manages timing, retries, and token budgeting. The frontend never touches the YouTube API - it only reads scored data.

Three specialized extraction agents work hierarchically, building context as they go:

Pass 1 - Context Agents: Analyze full text to detect assemblies and build hierarchical structure. Create a NumberSpace snapshot for subsequent passes.

Pass 2 - Component Agents: Extract components assembly-by-assembly, using Pass 1 context to avoid duplication and maintain naming consistency. Each component gets typed, tagged, and positioned in the hierarchy.

Pass 3 - Connection Agents: Detect physical connectors (bolts, cables) and manufacturing processes (welds, threads). Use exact hierarchical names from Pass 2 to prevent duplicate subject creation.

Human-in-the-Loop: After each extraction, agents present facts for human review. Users approve, edit, or reject before saving. Agents learn from corrections through citation tracking.

Progressive Refinement: The same bicycle can be described multiple times, adding detail incrementally. Agents check for existing subjects and extend them rather than duplicating.