CH 02
Nobody could defend the number until the system behind it was repaired.
Growth teams needed to know whether a shipped change worked. Answering that crossed acquisition onboarding, attribution, event pipelines and launch operations, and no one of them owned the answer. Five blocks: put the product on the platform, find where the meaning changed, repair it without discarding the run, then ship on it — and refuse a reported number until attribution is checkable.
Role spine
Senior Software Engineer, Growth · 2024–2025
Engineering Manager, Growth · 2020–2024
Senior Software Engineer, Growth · 2019–2020
The management period is experiment operations for teams running cross-product experimentation; the Atlassians in Mentoring platform was built concurrently.
An acquired product had to join the experimentation platform before it could be measured at all.
I led cross-functional work to onboard Atlassian’s largest acquisition onto experimentation and growth infrastructure — the prerequisite for any of the measurement work above applying to it.
Contract: a product is not part of the growth system until its events and attribution are on experimentation infrastructure.
10×
exceeded associated business OKR targets
Acquisition onboarding scope · associated business OKR
Cross Flow reported an uplift the prior approach could not defend.
A funnel observability audit established where the prior approach over-attributed credit. The attribution formula authored from that audit was subsequently adopted for Growth Experiment Impact Estimation, so the correction outlived the experiment that prompted it.
Contract: an uplift is reportable once its attribution is reproducible by formula, not by the dashboard that happened to render it.
>10%
prior-approach over-attribution exposed
Cross Flow funnel audit · formula subsequently adopted for Growth Experiment Impact Estimation
The same in-flight experiment said different things in different windows.
Analysis across StatSig attribution windows showed the reported uplift of experiments already running moving over a wide band. The range itself was the finding: no single window could be quoted as the result.
Contract: a result carries the window it was measured in, or it is not a result.
9%–41%
uplift variance across StatSig attribution windows
In-flight experiments · Statsig · the range is the finding
The Loom event pipeline stopped attributing while experiments were still running.
Embedded event-pipeline work salvaged attribution mid-flight: lost Cross-flow experiment data was recovered, roughly half of paid-user events that would otherwise have been excluded were preserved, and experiments that had been blocked could continue.
Contract: statistical validity is preserved by repairing the pipeline, not by restarting the experiment.
20%
lost Cross-flow experiment data recovered
Loom event pipeline · in-flight experiments
~50%
paid-user events preserved
Loom event pipeline · attribution salvage
5
in-flight experiments unblocked
Loom event pipeline · experiments already running
Three Cross Flow experiments shipped in Admin Hub without an incident or a restart.
Admin Hub experimentation covered global Loom requests, run as three Cross Flow experiments launched with zero incidents and zero restarts — the operational half of the same discipline.
Contract: an experiment that has to be restarted has already lost its result.
35%
D1D6AI increase
Admin Hub experimentation · three Cross Flow experiments · zero-incident, zero-restart launch
Artifacts and related writing
The same contract and qualified figure, stated as part of the home-page argument.
This work sits inside a private platform. Figures above cite the career inventory rather than a public URL.
The nearest published reasoning applies the same discipline — what the metric is allowed to count — on my own telemetry: Active players looked real until we asked which sessions counted →
Same method, live product — Codenames AI → · All field reports — articles →