ASG3-System

Inspiration

ASG-3 was born from the realization that most organizational knowledge is "fluid"—trapped in unindexed READMEs, wikis, and Slack threads—while data catalogs remain "rigid" and often empty. We aimed to solve the "cold-start" problem where newly ingested assets lack the semantic context necessary for effective governance. By creating an autonomous regulatory bridge, ASG-3 transforms unstructured tribal knowledge into verified, citable metadata facts within DataHub.

What it does

ASG-3 is a neuro-symbolic multi-agent system that automates data governance as a service. It executes a closed-loop Monitor–Generate–Reason (MGR) cycle:

• Monitor: Ingests declarative sources like GitHub and Confluence to find documentation gaps.
• Generate: Three specialists (Privacy, Domain, and Documenter) independently propose tags, glossary terms, and descriptions using a multi-vendor LLM panel.
• Reason: A symbolic consensus engine validates proposals, applying a mandatory Privacy veto to protect sensitive PHI/PII.
• Action: It calculates a risk-aware Importance Score via lineage analysis, automatically elevating the consensus threshold for critical assets. Approved changes are written back to DataHub via a deterministic Code Execution Pattern, leaving a permanent Reasoning Trail for auditors.

How we built it

The system follows a rigorous five-layer engineering archetype pinned to Python 3.11. Key components include:

• Orchestration: Powered by LangGraph to manage complex state transitions and human-in-the-loop (HITL) checkpoints.
• Context Management: Implements "The 3 Rs"—Relevance (RAG for glossary terms), Reliability (lineage-based importance), and Retention (few-shot Skill Library in ChromaDB).
• DataHub Integration: Leverages the Model Context Protocol (MCP) and the DataHub SDK for snapshots and write-backs.
• Design Patterns: Utilizes the Memento pattern for before/after snapshots, the Strategy pattern for interchangeable agent reasoning, and the Bulkhead pattern for fault tolerance.

Challenges we ran into

Implementing autonomous governance on a real clinical dataset of 55,500 records surfaced significant technical hurdles:

• Consensus Arithmetic: Standard quorums diluted the authority of single specialists (e.g., the Documenter for descriptions), requiring a refactor to use "competent panel" denominators.
• PII Accuracy: Standard symbolic scans failed on clinical schemas where hospital names were often misclassified as personal names, necessitating an EXACT_INDICATORS tier.
• Persistence Gaps: We discovered that the DataHub search index requires a full Tag entity lifecycle (upserting the entity, not just the link) for classifications to be searchable.
• Normalizing Agreement: Agents often agreed on categories but used different string orderings, which we solved by implementing set-based normalization for structured properties.

Accomplishments that we're proud of

• Risk-Aware Thresholds: We built a system where the consensus bar is earned, not asserted; mart_billing correctly triggered an 0.81 threshold based on its downstream impact on executive dashboards.
• Audit Rigor: Every metadata fact in the catalog now carries mandatory provenance, citing the exact line in a README or wiki that justifies its existence.
• Visual Governance: ASG-3 publishes Vega-Lite charts directly into DataHub, visualizing the "Before vs. After" catalog health and the internal debate logic of the agents.
• Code Execution Write-back: Consolidating multiple operations into a single deterministic script drastically reduced context window usage and eliminated redundant API round-trips.

What we learned

Tribal knowledge is frequently contradictory; we learned that effective governance is not just about retrieval, but about adjudication between competing claims (e.g., an obsolete README vs. an updated Confluence page). We also found that multi-vendor model panels (GPT, Claude, Gemini) are essential for true consensus reliability, as a single vendor tends to share the same underlying biases across different prompts.

What's next for ASG3-System

The roadmap for ASG-3 includes:

• Industry Expansion: Scaling the symbolic indicator engine to support financial, legal, and manufacturing domains beyond healthcare.
• Deep Governance: Implementing full support for Domains and complex ownership hierarchies.
• Enhanced Perception: Moving from substring matching to semantic chunking for large technical manuals to further optimize token efficiency.
• Self-Evolution: Enabling the system to autonomously extend its own toolset when encountering unknown metadata types.

Built With

• 3-rs
• asg-3
• autonomous-semantic-governance
• code-execution-pattern
• consensus-engine
• context-management
• data-lineage
• datahub
• governance-as-a-service
• hitl
• importance-score
• langgraph
• mcp
• memento-pattern
• mgr-cycle
• multi-agent
• neuro-symbolic
• pii/phi-detection
• privacy-veto
• python-3.11
• rag
• reasoning-trail
• structured-properties
• tdd