AI Cut Our Research Time From 7 Days to 15 Minutes

Gareth Yoder, Vanguard15:39 · Oct 2026 · 1,051 views
Thumbnail for AI Cut Our Research Time From 7 Days to 15 Minutes Watch on YouTube
TL;DR
  1. 1

    Vanguard modernized a cash ledger handling up to 30 million calls per day by running data, producer, and consumer migrations in parallel.

  2. 2

    Exposing the legacy codebase through an MCP server reduced research work for mainframe engineers from seven days to 15 minutes.

  3. 3

    Migration documentation, incremental AI adoption, rethinking the target design, and observability were all treated as part of the modernization work.

Summary

Gareth Yoder describes Vanguard's modernization of a cash ledger that handles money movement across brokerage accounts, supports web, mobile, and batch channels, and receives up to 30 million calls per day. The team ran data, producer, and consumer migrations in parallel. Early research depended on mainframe engineers and architects, but exposing the legacy codebase to AI tools through an MCP server reduced one analysis cycle from seven days to 15 minutes. Vanguard used change data capture, row and column audits, an API gateway, idempotent operations, schema registries, GraphQL, data shares, and Kafka events during the migration. Yoder also explains how business-use-case documentation improved consumer onboarding and incident response. His advice is practical: make legacy systems queryable by AI, treat migration documentation as product material, redesign instead of copying the old system, adopt AI as it becomes useful, and keep observability in place throughout the transition.

Key ideas
00:57

The cash ledger is a high-volume system with several access patterns

Yoder defines the cash ledger as the system for money movement into Vanguard brokerage accounts and between accounts. It supports payment rails such as ACH and wires, along with web, mobile, and batch channels. The system runs in an active-active, multi-region, fault-tolerant setup. It receives up to 30 million calls per day, while clients generate about half a million ledgers daily. Vanguard split the modernization into data, producer, and consumer migrations, then ran all three at the same time because modernization was an enterprise priority.

02:09

An MCP server turned legacy-code research into a 15-minute feedback loop

At the start of the work, mainframe engineers and architects had to inspect the legacy code and answer questions about the data schema. A turning point came when Vanguard made the legacy codebase available to AI tooling through an MCP server. Analysis that had taken a mainframe engineer seven days could then be done through an AI prompt. Yoder says the feedback loop fell to 15 minutes, which helped product owners, technical leads, and developers work through modernization questions faster.

03:15

The data migration used simple checks instead of an elaborate temporary system

Vanguard used change data capture and scanned both the legacy and modern databases. The team ran daily audits. It checked row-level counts for consistency, then compared individual columns for integrity. Amounts, account information, state withholding, federal withholding, and other attributes had to match between the legacy database and the cloud system. Yoder warns against over-engineering this migration code because the code would be deprecated after modernization was complete.

04:06

The producer migration replaced hidden database writes with a governed gateway

Legacy producers reached the infrastructure through service requests, APIs that wrapped those requests, and direct table inserts. Direct inserts made it difficult to know which system had created data. Vanguard put a common data schema behind a gateway that handled cloud data inserts. The redesigned producer system added idempotency and consistency across multiple writers, using an approach similar to a Snowflake ID. Yoder describes this as a redesign rather than a lift-and-shift migration.

05:17

AI-accessible documentation reduced support work for 35 producers

Thirty-five producers needed to move to the new ecosystem. Vanguard first wrote manual migration documentation, then stored it in the company wiki and exposed it through the MCP server. Producers could use prompts instead of holding repeated support chats with the modernization team. Yoder says the number of chats and the amount of producer support work fell after the MCP became available. AI tools also helped translate legacy status codes into human-readable ledger enums in the cloud system.

06:30

Consumer access was separated across real-time, batch, and streaming paths

Vanguard supported three consumer patterns: real-time access, batch access, and streaming. A schema registry provided a common structure across the channels. Real-time consumers used GraphQL and subscribed only to the attributes they needed. Batch consumers received data-share views from the warehouse, with each view tailored to its use case. Kafka topics streamed database inserts and updates to teams such as fraud. At-least-once delivery, combined with idempotent operations, let consumers retry failed work without creating duplicate ledgers.

08:52

Business descriptions made consumer inventory useful during incidents

Vanguard had 50 consumers that needed ledger data. Each application had to register with a three-character application code, but the team also required a description and business justification for the data request. Examples included a trade order fulfilling a cash obligation, a retirement account making a required distribution, and a 529 account receiving a contribution. AI tooling could compile these descriptions into an inventory of modernized use cases and business functions. The same information could help responders judge whether a failing application had hours or only minutes before its service-level impact became serious.

11:03

The modernization program kept a shared team brain and fixed forward

Yoder says 90% of the cash program's engineers use AI tools daily. Teams use multi-agent workflows and shared knowledge bases. Their idea of a team brain stores product knowledge so a new engineer can prompt against the same context as someone who has worked on the system for five years. While both codebases exist, Yoder advises teams to fix the modernized state instead of repairing decades-old defects in the legacy system. That approach can encourage producers and consumers to move to the cloud, but it requires allowlisting because audit logs will change.

13:14

The five lessons put migration material and observability alongside code

Yoder's five lessons are to make legacy code AI-queryable, treat migration documentation as part of the product, reimagine the target instead of lifting and shifting, adopt AI incrementally, and treat observability as non-negotiable. Migration documentation matters because producers depend on it while onboarding, even if it is eventually retired. Teams should also track who produces and consumes data, with metrics, alarms, and traceability in place before the modernized system becomes normal operations.

"Now that communication and that back and forth could all be done via the prompt on the producer side."05:42
Who should watch
  • You are moving a high-volume legacy system while producers and consumers still depend on its old interfaces.
  • Your migration team spends days answering questions about undocumented mainframe code or repeatedly supporting application teams.
  • You need a concrete way to connect AI-assisted code research with data validation, onboarding, incident response, and observability.