Bridging 1995 Mainframes with Agentic AI Built for Google Fragfest Hackathon - Track B
Decades of accumulated process debt in the enterprise sector. Most AI pilots fail (95%) not due to model capability, but the inability to safely integrate with ancient mainframes, siloed databases, compliance layers, and the tribal knowledge held by retiring experts.
- India’s $300B+ IT services industry is at risk of losing its global edge unless it masters legacy integration in the agentic AI era.
- We don’t have an AI problem. We have a Legacy + Tribal Knowledge + Compliance Debt problem.
LegacyGuard AI is a multi-agent system that acts as a "Digital Archaeologist + Safe Executor" for legacy environments. Instead of feeding a raw database dump to a single hallucination-prone LLM prompt, we built a tiered, deterministic orchestration pipeline.
- Scout Agent (gemini-3.1-flash-lite): Ingests raw, undocumented fixed-width strings (e.g., COBOL/EBCDIC) and dynamically maps the positional schema.
- Knowledge Bridge Agent (gemini-3.1-flash-lite): Applies human "tribal knowledge" to translate raw mappings and codes into actionable, plain-English profiles.
- Compliance Agent (gemini-3.5-flash): Acts as the strict guardrail, evaluating the modernized profile against rigid regulatory frameworks (e.g., RBI Guidelines) to approve or halt workflows.
- Execution Orchestrator (Boss Agent): Coordinates the Python flow, handles API failovers (503 traffic spikes) gracefully, and logs every step into an immutable Audit Trail.
To prove our orchestration works, we simulate a legacy banking core using a mock VSAM hierarchical file structure. When queried, it does not return clean JSON; it returns an undocumented, 46-character fixed-width flat file (COBOL Copybook format).
Available Test Cases:
10045892: Clean / Medium Risk (Demonstrates auto-approval)10045893: Closed / High Risk (Demonstrates Compliance Agent halting the transaction)99999999: Custom Test (Demonstrates dynamic lookup)
Because this system requires direct access to enterprise VPCs and sensitive data, the MVP is designed to be run locally, simulating an internal enterprise deployment.
1. Clone the repository:
git clone (https://github.com/JensonCode007/Multi-Agent-Orchestrator.git/)