Legal AI Feature Delivery Speed Doubled

Microservices architecture with Orchestrator agent that receives user requests and routes them to proper skill agents. Core MCP service acts as a shared tool and context registry, enabling any agent to access approved external data sources through a standardized interface. LangChain for LLM chain construction; LangGraph for stateful agent workflow orchestration.

Pynest built the transaction processing flow end-to-end, from intake through validation, pricing reconciliation, discrepancy analysis and upload. The team designed a 5-step validation process for Excel spreadsheets of varying quality, with a pre-upload stage that stops invalid data from entering the core system. A structured discrepancy report process runs with the advertiser clients and tracks each resolution. The full backlog of 12,000+ transactions was batch-uploaded into the core system, and the team documented the work as script specifications so it can be automated later.

Services are containerized and deployed inside the bank internal infrastructure, through the existing CI/CD pipeline rather than a new one. Aligning with the bank's own tooling kept the migration inside its security perimeter and avoided new dependencies in a regulated environment. MCP was chosen as the integration layer because it cuts the ongoing cost of writing custom third-party integration code.

Every service was adapted to meet the bank internal security requirements before it shipped, and each one passed a formal security review first. Role-based access control governs which legal team users reach which capabilities. Data handling follows the bank data governance policies throughout. No data leaves the system outside the approved MCP integrations, so external egress stays inside a controlled, reviewed boundary.