GraphRAG Knowledge Platform over Legacy Mainframe

Context

A leading global insurance provider ran critical business logic on PL/1 mainframe systems built over decades. The pool of PL/1 specialists was shrinking, creating acute key-person risk. Batch flows and scheduler dependencies could not be traced end to end, business logic was buried in code with no mapping to process, and German-language table and column names steepened the learning curve — onboarding took months and every impact analysis was a manual effort.

Approach

• Designed a generative-AI analysis and knowledge-access platform on a RAG/GraphRAG architecture.
• Ran entirely on the client's internal LLM and embedding models to satisfy data-sovereignty constraints — no external inference.
• Analysed PL/1 source, JCL flows, batch schedules, and database structures, using a graph layer to give fragmented assets semantic meaning and traceable relationships including domain values.
• Made batch flows, dependencies, and database relationships visually navigable for developers and business analysts alike.
• Specified AI agents and tooling for platform monitoring, logging, and management-level usage reporting.

Outcome

• Adopted by 70+ developers and analysts.
• Manual, effort-intensive impact analysis replaced by on-demand dependency tracing.
• Reduced dependency on scarce PL/1 specialists, mitigating key-person risk.
• Stronger analysis quality and traceability — end-to-end visibility improved the accuracy and auditability of change assessments.

Stack