Xero

Vice President / General Manager, Data. January 2021 to July 2024.

I joined Xero to lead an eight-person data enrichment team, covering models, transformations, and the sourcing and landing of data. Platform, architecture, product analytics, and enablement had separate leaders, so getting people onto shared tooling meant working across teams I did not manage.

My remit later expanded to the combined engineering, platform, architecture, product analytics, and enablement organization, with 65 people and a combined-organization budget of about $20 million; my peer led science. The expansion included consolidation of existing functions, not 57 people hired into the original enrichment team. The work brought together shared modeling, an ingestion replacement, product instrumentation, and the organization needed to deliver and support them.

Self-service modeling

Snowflake, dbt Core, SQL, Data Vault, Kubernetes, Terraform, Prefect, DataHub, WhereScape RED

Product teams could not wait in a data-team queue for every model they needed. With the engineering team I selected dbt Core, building on the existing Data Vault estate with conventions and reusable tooling that let product teams develop models in SQL. This expanded what they could do beyond Snowflake-native transformations and the existing WhereScape RED tooling. By July 2024, dbt was the standard for new modeling and WhereScape was treated as legacy.

Revenue Operations and about 20-25 product analysts became our first customers and development partners. They had already built tooling to get their work done, so there was useful experience in what they had made, as well as frustration with what the data team had provided. We combined their tooling with the engineering team's framework and worked toward something both groups could support. It took about 18 months for them to be using it and standing behind it.

YAML in each model's Git repository described ownership, schedules, and dependencies. GitHub Actions promoted that configuration into Prefect. DataHub made ownership and discovery visible, so product teams could find what was already there without unrestricted control of the infrastructure.

When people don't adopt a platform, I first ask whether the platform understands their work. The people who kept the function running get a clear account of what improves, what gets harder, and how they'll be supported. A central team's architectural preference does not settle those questions.

Product analytics

Custom frontend instrumentation, .NET telemetry libraries, Protocol Buffers, Kafka, Snowflake, dbt, Mixpanel, Salesforce

The existing frontend instrumentation showed user interactions, but understanding a product also required visibility into scheduled processes and actions taken by other people on a customer's behalf. I worked with the team and product leaders on modularizing the library for reuse across frontend interfaces and extending coverage into backend services through .NET telemetry libraries. Events described the product domain, the human or system actor, and the action, making background activity and work by accountants visible alongside the customer's own interactions.

For product analytics, I defined a migration from the inherited AWS Lambda, S3, and DynamoDB stack toward shared Kafka infrastructure and Kafka-to-Snowflake ingestion. I supported the team's selection of Protocol Buffers for those events and set latency requirements around the downstream analytical uses.

Mixpanel-side joins were costly and unreliable, so I directed enrichment into Snowflake, combining product events with customer, billing, regional, and marketing attributes before sending them to Mixpanel. The new path gave analysts more customer context and was running in parallel with the existing joins at handover, while adoption continued.

CDC architecture

Go, Kafka, SQL Server CDC, PostgreSQL, DynamoDB, Snowflake, Apache Iceberg, Amazon S3

The database-capture work addressed bespoke extraction across roughly 800 data sources, including about 600 SQL Server shards. Fivetran made useful data available while the teams developed a replacement for the high-volume change traffic. I led an evaluation of competing designs and proposed repeatable Go CDC source agents that would publish changes to Kafka, with transformation and modeling kept downstream.

Customers and their accounting practices could move between shards, so the design carried metadata about that movement with the data. That allowed downstream consumers to follow the customer without implementing the source system's migration logic themselves. The destinations were Apache Iceberg on S3 for historical data and Snowflake ingestion into a source Data Vault, with PostgreSQL and DynamoDB also covered by the design.

I recruited the principal engineer who had led the earlier Ticketmaster Snowflake implementation to lead that part of the work. The engineers took the architecture into implementation, which continued beyond my handover.

Science and executive planning

I partnered with the science leader on the direction for an operational data layer supporting analytics and intelligent products. The aim was to make source data and data enriched with business logic available to more consumers without each of them extracting it again from transactional applications. Snowflake would support exploration of source data and business rules, while history in Iceberg would support larger-scale processing and training.

Joining the senior leadership team brought that work into company-level planning with the CEO, CPTO, and Finance. It meant connecting choices about the data platform to product priorities, the company's AI plans, and the people and funding needed to carry the work through.

Leadership, delivery capacity, and succession

During the restructure in 2023, the science lead and I continued leading the work for about six months before the leadership gap above us was filled. We kept developing the shared data foundations and roadmap while the organization changed around the teams.

I also worked with managers on a new performance framework. Decisions about performance affect people's careers, so the framework needed to be understandable to the people being assessed and useful in conversations about expectations and development. The work included helping managers apply it and advocating for people affected by the change.

I established dedicated tiger teams for modernization so pulling an engineer onto other work required an explicit decision about the effect on delivery. A three-year Fivetran agreement with declining usage gave the replacement a timetable, with vendor capacity available while the team built the internal path. We reduced legacy processing and support as work moved across, spending less time repairing old systems while providing a supported destination for their users.

I developed a successor and stayed through the leadership transition before handing over in July 2024, with teams responsible for the roadmap and its continued delivery.

When a team's capacity is moved, I revisit the date or the scope as a leadership decision, so the contradiction doesn't land on the people least able to change it. When I hand over a program, I want my successor and the teams to inherit the authority and support to make those decisions, as well as the architecture and roadmap.

Delivery at handover and beyond

By handover, Fivetran had been ingesting for about a year, dbt was the standard for new modeling, and selected models were feeding executive dashboards. The wider Revenue Operations replacement was still in progress, Salesforce integration had started, and I had selected and negotiated the new BI tool.

The Snowflake-enriched Mixpanel path was running alongside the existing joins while teams moved onto the enriched datasets. The replacement CDC platform was at a different stage: its Go/Kafka architecture, Iceberg targets, roadmap, and delivery teams were established, but the new CDC-to-S3 path was not operating when I left.

Data refresh had moved from about half a day toward one to two hours, ahead of the six to eight hour business need. That gave us room to support faster uses of the data without committing the team to a service level the business had not asked for.

The team continued the program and later reported that the CDC replacement and planned Fivetran retirement landed around late 2025, followed by the broader model migration around February 2026.

Who did what

I led the architectural evaluation and contributed example code and limited modeling, alongside the organizational work needed to fund and protect delivery. The engineers built the production pipelines and models, and my successor and the teams carried the replacement through after I left.