The household intelligence layer, rebuilt as a self-serve platform

Outra sells household intelligence: the real-world context that sits underneath a marketing database. Moving patterns, life stage, affluence, tenure. Coverage runs to more than thirty million UK households, and the data itself was never the problem.

Getting it to customers was. Enrichment ran as a service: a client sent a file, an analyst matched it, cleaned it, and sent it back. Every engagement consumed analyst time that did not scale, and the turnaround made the data feel like a research project rather than part of a marketing stack. Audience discovery, the higher-value half of the proposition, barely reached customers at all because nobody had the time to run it.

What we did

We took it as a managed team with accountability for a defined outcome: the same enrichment and discovery, driven by the customer, in a browser, at platform.outra.co.uk.

1. Scoping, one week. Two engineers and a delivery lead sat with the analyst team and watched them work. The specification came out of what they actually did, not out of a requirements document. Most of the value was in the steps they had stopped noticing they performed.
2. Factory setup, week one. Environments, CI, the delivery platform and a masked slice of production data stood up against their GCP project before build started.
3. Build, weeks two to eleven. Five certified engineers, shipping weekly. Agents took the matching and transformation logic that existed as analyst SQL and turned it into tested dbt models; the engineers held the identity resolution and the privacy boundary, which is where this class of system goes wrong.
4. Hardening and handover, weeks twelve to fourteen. Load testing against full-file uploads, audit logging for their SOC 2 and ISO 27001 obligations, runbooks, and shadow support through the first customer cohort.

What made the difference

The weekly cadence mattered more than usual here, because the analysts were the domain experts and the users at the same time. They saw working software every Thursday from week two. Roughly a third of what shipped was not in the original scope, and it displaced things that were, because eleven weeks of their feedback beat a specification written in week one.

The second saving was refusing to rebuild the matching engine. It worked. It was slow and undocumented, but correct, and correctness in identity resolution is expensive to re-earn. We wrapped it, tested it properly for the first time, and spent the budget on the platform around it instead.

Where it landed

Enrichment and audience discovery now run self-serve, with weekly audience refreshes and activation into the channels customers already buy. The analyst team moved off file-shuffling and onto the modelling work that actually needs them.