How complex questions are investigated

Context

An airport out-of-home campaign was intended to prompt existing users to record upcoming travel. Individual ad exposure was not observable and a randomised control group was unavailable.

What I did

1. Built a transparent exposure proxy from airport-area app activity followed by activity abroad, while excluding local residents and nearby overland destinations.
2. Defined pre-campaign and campaign windows and compared likely exposed users with a non-airport control population across multiple years.
3. Separated reach from conversion and segmented users by tenure and prior engagement to test whether changes were behavioural or compositional.
4. Ran sensitivity checks with alternative attribution windows, an excluded pull-forward period, other airports, and reweighted user mixes.

Context

A temporary pricing campaign offered business customers an additional personal payment product through several email waves. The analysis needed to separate message sequence, pricing, and natural purchase timing.

What I did

1. Linked masked customer, product, and campaign histories in SQL and reconstructed the paths created by launch, reminder, and follow-up communications.
2. Built acquisition cohorts with a standardised twelve-month observation period to compare product adoption and purchase timing across years.
3. Modelled time to purchase in R with right-censored observations, Kaplan-Meier curves, a log-rank test, and a Cox proportional-hazards model.
4. Built a pre-campaign Poisson baseline to compare expected and observed purchase patterns across both recent and established cohorts.

Context

Market share among newly registered sole proprietors could only be observed after entrepreneurs opened a business account. That delay made recent months incomplete and unsuitable for direct comparison with mature cohorts.

What I did

1. Matched external company registrations to internal open and closed account records using unique business identifiers and aligned eligibility rules across numerator and denominator.
2. Analysed the delay between company registration and account opening and defined a fixed sixty-day maturation window with intermediate measurement points.
3. Tested whether early conversion could predict the mature market-share estimate and used the model to nowcast recent, incomplete cohorts.
4. Designed a reporting approach that separates mature observations, forecasts, late conversions, pending applications, and the early indicator used for day-to-day steering.

Context

Portfolio information was spread across statements, fund documents, spreadsheets, and live market sources. Private-market cash flows and inconsistent document formats made a single current view difficult to maintain.

What I did

1. Designed a shared data model for listed assets, private-equity funds, capital calls, distributions, book values, market values, and allocation targets.
2. Built an extraction pipeline that uses OCR and language models to turn statements and fund documents into structured records with source references.
3. Combined extracted records with live market data and explicit data-quality checks so uncertain or conflicting values remain visible for review.
4. Delivered the workflow as a production web product, covering ingestion, validation, portfolio monitoring, and decision-ready views.

Context

A small data team faced growing demand from clients, operations, product, and commercial teams. Reporting was reactive, ownership was unclear, and the technical foundation needed to mature without slowing day-to-day delivery.

What I did

1. Defined the function's mandate, translated competing requests into a shared roadmap, and introduced clear ownership and prioritisation.
2. Designed the team shape and hiring sequence across analytics, engineering, and data science, while coaching people into broader responsibilities.
3. Introduced layered data models, version control, review practices, shared definitions, and data-quality checks as the basis for reliable delivery.
4. Replaced fragmented reporting with reusable client and operational products, while creating a regular working rhythm with product, commercial, and operations leaders.