AI Cut Our Research Time From 7 Days to 15 Minutes — Gareth Yoder, Vanguard

Highlights

• AI integration via an MCP server reduced legacy code research cycles from seven days down to 15 minutes.

• The Vanguard Cache Program receives up to 30 million calls per day and generates half a million ledgers daily.

• Ninety percent of engineers in the Cache Program utilize AI tools on a daily basis.

• Data migration verification utilized daily change data capture audits checking row-level counts and column integrity.

• Producers connected to the legacy system through 35 distinct channels including service requests, wrapper APIs, and direct table inserts.

• Modernization eliminated technical debt by abandoning lift-and-shift strategies in favor of reimagined fintech concepts like idempotency.

Timeline

1. 00:00 - 01:52

   Cash Ledger Ecosystem and Modernization Pillars

   Vanguard launched its cash plus offering for brokerage clients in 2022, initiating a modernization pipeline for legacy codebases. The ledger ecosystem generates half a million ledgers every day while operating in a fault-tolerant environment. Executing the enterprise modernization priority demanded running all three migration pillars in parallel.

   • The Vanguard brokerage account cash ledger processes ACH, wires, and payment rails across web, mobile, and batch channels.
   • System infrastructure runs in an active, active, multi-region ecosystem handling 30 million calls per day.
   • Modernization required executing data migration, producer migration, and consumer migration simultaneously.
2. 01:52 - 03:20

   AI Tooling and Research Analysis Transformation

   Before AI tooling availability, determining data schema models required lengthy manual reviews by mainframe engineers. Introducing an MCP server that exposed the legacy codebase directly to AI models transformed the workflow. This technological shift enabled product owners, technical leads, and developers to obtain instant answers.

   • Initial research relied on manual consultations with mainframe engineers and enterprise architects to map data schemas.
   • Connecting the legacy codebase to AI tooling via an MCP server established faster feedback loops.
   • Analysis turnaround time dropped from seven days down to 15 minutes through direct AI prompt execution.
3. 03:20 - 06:38

   Data Migration and Producer Onboarding

   Data validation relied on strong consistency checks for row-level counts and precise integrity matching for financial attributes like state and federal withholding. Preventing over-engineering kept temporary migration code clean since old systems would be deprecated post-migration. Producer onboarding utilized shared wiki documentation exposed through AI chat prompts to streamline integration.

   • Change data capture approaches scanned databases daily to verify row counts and column attributes between legacy and cloud systems.
   • Producer access relied on service requests, wrapper APIs, and direct table inserts into the database.
   • Exposing migration documentation via an MCP server reduced the volume of direct support chats required by engineering teams.
4. 06:38 - 11:12

   Consumer Migration Channels and Business Use Cases

   Real-time balance checks utilized GraphQL endpoints while Kafka event topics streamed inserts and updates for fraud detection teams. At-least-once delivery models integrated idempotency safeguards to prevent duplicate ledgers during partial job failures. Compiling business descriptions into AI tools allowed engineering teams to assess availability impact during incidents.

   • Consumer data access operated across real-time, batch, and event-streaming paradigms backed by a unified schema registry.
   • GraphQL powered real-time GraphQL API calls for account balance displays, while data shares served warehouse analytics.
   • Consumers submitted explicit business justifications and application codes to map functional dependencies and incident SLAs.
5. 11:12 - 15:29

   Key Leadership Lessons in Modernization

   Modernization requires avoiding direct lift-and-shift strategies in favor of reimagining ecosystems around modern capabilities. Observability remains non-negotiable for identifying data producers and consumers during daily business operations. Making legacy code queryable via AI and maintaining strict metric alarms secures long-term operational success.

   • Ninety percent of engineers utilize AI tools daily, relying on shared team knowledge bases and multi-agent workflows.
   • Fixing forward rather than patching legacy code provides incentives for producers and consumers to adopt the cloud.
   • Treating migration documentation as a first-class product ensures smooth producer onboarding during transitional phases.