Gareth Yoder, cash program technical lead at Vanguard, describes a cash-ledger modernization on AI Engineer. A 15-minute talk.

Make legacy knowledge available to the tools

  • Yoder reports up to 30 million service calls and about half a million new ledgers daily, across multiple active regions.
  • Data, producer and consumer migrations had to proceed in parallel.
  • Exposing legacy code through MCP, a standard for connecting AI tools to systems, shortened a research feedback loop from seven days to 15 minutes.
  • Migration documentation went through the same connection, reducing repeated support conversations with 35 producer teams.

Keep the migration verifiable

  • Change data capture copies database changes; daily audits compared row counts and individual fields between legacy and cloud databases.
  • Checks covered money amounts, account information and tax withholding; temporary comparison code stayed simple.
  • A common gateway replaced a mix of service requests, wrapped APIs and difficult-to-trace direct table inserts.
  • Idempotent operations make retries safe by preventing duplicate effects, including duplicate ledger creation.

Give consumers the data their job requires

  • A shared schema supports real-time GraphQL requests, tailored warehouse views and Kafka event streams.
  • At-least-once event delivery can repeat messages, so consumers need safe duplicate handling.
  • Fifty consumers supplied business justifications alongside application identifiers, making their dependencies easier to understand.
  • Connecting those descriptions to logs and traces can help distinguish an urgent customer-facing failure from a batch delay.

Preserve shared context and operational ownership

  • Yoder says 90% of program engineers use AI tools daily; a shared “team brain” makes accumulated knowledge available to newcomers.
  • Improve the modern system as old bugs emerge, recording intentional differences so reconciliation checks remain meaningful.
  • Treat migration docs as part of the product, adopt AI incrementally and retain metrics, alarms and traceability after the migration.

Caveat: seven days to 15 minutes describes Yoder’s reported research feedback loop. The talk supplies neither a controlled productivity study nor evidence that the entire migration became that much faster.

“Observability is non-negotiable.” — Gareth Yoder