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

Loading slides…

Read the talk

AI Cut Our Research Time From 7 Days to 15 Minutes

Yoder reports the research feedback loop falling from seven days to 15 minutes.Open full source frame
Yoder reports the research feedback loop falling from seven days to 15 minutes.

Gareth Yoder explains how Vanguard modernized a cash ledger through parallel data, producer, and consumer migrations—and made its legacy code and migration guidance available to AI tools.

From a talk by Gareth Yoder

At a glance

Ideas worth remembering

  • Exposing legacy code through MCP let the team ask AI tools research questions directly; Yoder reports a feedback loop falling from seven days to 15 minutes.

  • Daily migration audits compared both row counts and individual attributes. Expected differences from fixes made only in the modern system needed allowlisting.

  • A common gateway gave 35 producers a modern write path, while queryable migration documentation reduced the team’s reported support burden.

  • Real-time, batch, and streaming consumers had distinct access patterns; at-least-once event delivery made safe retry handling important.

  • Consumer business justifications can connect an application’s technical failure to its impact, while metrics, alarms, and traces remain essential for daily operations.

A ledger to modernize while it keeps moving

Vanguard’s cash ledger tracks money moving into brokerage accounts and between accounts, across payment rails such as ACH and wires. It serves web, mobile, and batch channels in an active-active, multi-region system. Gareth Yoder, the cash program’s technical lead, says its services receive up to 30 million calls a day and clients generate half a million ledgers daily.

He introduces the three parallel migration pillars.Open full source frame
He introduces the three parallel migration pillars.

After the 2022 launch of Cash Plus, the program took on modernization as an enterprise priority. Data, producers that write it, and consumers that read it all had to migrate in parallel. That made understanding the existing system an immediate dependency: engineers and architects needed to work out what the new data schema should contain.

Initially, answering a legacy-code question meant asking a mainframe engineer to investigate and return with an analysis. Once the codebase became available to AI tools through an MCP server, the team could prompt the tool directly. Yoder reports that a research feedback loop fell from seven days to 15 minutes. The comparison describes the team’s experience with that research loop; it does not establish that every answer or migration task took 15 minutes.

0:120:42
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:12 · section reference included

Move the data and check the details

The data migration used change data capture (CDC), followed by daily scans of the legacy and cloud databases. The team compared row counts, then checked individual attributes: amounts, account information, and state and federal withholding. The distinction matters: matching counts can reveal missing or extra rows, while column comparisons can reveal records whose contents differ.

He describes checking individual attributes across the legacy and cloud databases.Open full source frame
He describes checking individual attributes across the legacy and cloud databases.

Yoder cautions against over-engineering the audit solution. Its job was to support the transition, and the code would be deprecated when modernization was complete. The checks provided migration evidence; matching counts alone would not establish that every value matched.

3:113:41
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

3:11 · section reference included

Make writes visible and give producers a migration path

How did the old write paths compare with the new one? Legacy producers used a service request, an API wrapping that request, or direct table inserts. Direct inserts were particularly troublesome: data appeared in the database without the team necessarily knowing who had produced it. The new design placed a common data schema on a gateway wrapping the APIs that insert cloud data, giving producers a defined route into the modern system.

He describes the gateway wrapping the cloud insert APIs, immediately after explaining the legacy write paths.Open full source frame
He describes the gateway wrapping the cloud insert APIs, immediately after explaining the legacy write paths.

The move also let the team change behavior. It adopted idempotency—the property that a repeated operation can be handled without creating a second effect—and sought strong consistency across multiple writers. Yoder describes using a paradigm similar to a Snowflake ID, though he does not specify its implementation. The goal was to redesign the producer offering around modern capabilities rather than reproduce every legacy path.

There were 35 producers to bring across. The team first wrote migration documentation manually, published it in the company wiki, and later exposed it through MCP. Producers could then ask questions through prompts instead of relying as heavily on separate support chats. Yoder reports a large reduction in those chats and support requests.

The cloud ledger also used human-readable enum values where the legacy system used status codes. Yoder says the AI tools helped bridge the intent between the old and new representations. The migration guidance was therefore part of the product during onboarding, even though it too would eventually be retired.

4:124:42
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:12 · section reference included

Three ways to consume the ledger

Consumers had different timing and data needs, so the modern system offered three access patterns with a schema registry across the channels:

He introduces real-time, batch, and streaming access with a common schema registry.Open full source frame
He introduces real-time, batch, and streaming access with a common schema registry.
  • Real time: The heavily used funds-available API helps determine the balance displayed on an account. Consumers call through GraphQL and select the attributes their use cases need.
  • Batch: For reconciliation or analytics, the data warehouse exposes data-share views tailored to each consumer’s stated needs. Data governance supported limiting each consumer to the attributes it needs.
  • Streaming: A Kafka topic carries database inserts and updates. Fraud teams can ingest those events into their tooling in real time.

The event stream provides at-least-once delivery: a consumer can receive an event more than once. That makes idempotent handling important when work is retried. Yoder gives the example of a batch job failing halfway through; a retry needs to avoid creating duplicate ledgers while the event stream still delivers the ledger events. He does not detail the implementation of those safeguards.

6:387:08
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

6:38 · section reference included

Use onboarding context during incidents

Fifty consumers needed the ledger. Alongside Vanguard’s three-character application code, the team asked each one for a description and business justification. A trade order might need cash data to meet its obligation; a retirement application might support a required distribution; a 529 application might support a contribution. Those descriptions say what the data enables, not just which application receives it.

He discusses using consumer business descriptions to understand supported functions and incident impact.Open full source frame
He discusses using consumer business descriptions to understand supported functions and incident impact.

AI tools can compile those descriptions into a view of modernized use cases and the business functions the ledger supports. They also make an incident question easier to ask: what happens if a service-level agreement is breached? Yoder contrasts a batch application, where the team may have hours to respond, with a real-time API returning errors to a client, where it may have minutes.

The team was still evaluating how to combine business justifications with observability and log traces to improve response time. The useful link is between an application identifier in an incident and the human consequence of its failure; the integration was presented as ongoing work.

8:589:29
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

8:58 · section reference included

Fix forward, share context, and keep observability

AI arrived after the migration had begun, and the teams adopted it as their tools changed. Yoder reports that 90% of cash-program engineers now use AI tools daily for work that includes research, development, and testing. Teams also use multi-agent workflows and shared knowledge bases. In the “team brain” approach, knowledge generated while building features goes into common storage, where a new engineer can bring it into a prompt alongside context available to long-standing teammates.

He discusses daily AI use, multi-agent workflows, and shared knowledge bases.Open full source frame
He discusses daily AI use, multi-agent workflows, and shared knowledge bases.

Maintaining two codebases creates a harder choice when the team discovers an old bug. Yoder recommends fixing it in the modern system and using that improvement as an incentive to migrate. That can make the databases intentionally differ, so the audit process needs a way to allowlist expected discrepancies rather than treat every difference as a migration failure.

The most difficult producers and consumers to onboard, in Yoder’s experience, were those trying to transfer the old design directly. Teams that reconsidered their products could use the modern ledger’s capabilities. His closing advice joins that redesign with practical support: make legacy code queryable, maintain migration docs while people need them, and adopt AI incrementally as tools become available.

Observability remains necessary after the migration. Metrics, alarms, and traces help teams see who produces and consumes data and support the ledger in daily operations. The direct inserts that once appeared without an identifiable producer show why that visibility matters.

10:5911:29
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:59 · section reference included

Resources

Transcript download
Full transcript

Loading transcript…