Customer story

Pulse Telecom

Churn modelling at the speed of product, not the speed of legal.

Pulse's growth team replaced their slow, painfully governed access to subscriber data with a daily synthetic refresh. Experiments per quarter went from 6 to 41.

At a glance

Industry
Telecom
HQ
Lisbon, Portugal
Size
5,800 employees · 14M subscribers
Stack
Doppelset Lab, BigQuery · dbt

6.8×

more churn experiments per quarter

−61%

reduction in real-data access tickets

4 weeks

shaved from each model release

The challenge

Pulse's churn-modelling pipeline depended on a curated slice of subscriber data: tenure, plan, usage, support contacts. To touch it, data scientists had to file an access ticket, justify a use case, and wait an average of nine working days for legal sign-off — a process that effectively capped how many ideas the team could test each quarter.

When marketing wanted to A/B-test a retention offer, the team couldn't share even a sampled view of the subscriber base with the agency that designed the experiment. Pulse was losing weeks on every campaign to the back-and-forth.

The solution

Pulse moved their churn data into a Doppelset Lab tenant. A daily job retrains a doppel of the subscriber base with ε = 1.0 and rebalances the churn label to 25% so models have enough positive signal during prototyping.

Data scientists now point their notebooks at the synthetic churn warehouse by default. When a model graduates from prototype to candidate, an automated job re-runs the final training against production data in a sealed environment — and bakes a quality + privacy receipt into the model card.

Marketing and their external agency share the same synthetic dataset for campaign design. The first 18 hours of every campaign — including segment definition and creative testing — now happen entirely on synthetic data.

Doppelset turned a 'no' from legal into a 'yes, by Friday'. Their differential-privacy report is the cleanest I've ever sent to a regulator.
Priya Nair, Chief Data Officer, Pulse Telecom

The results

From 6 to 41 experiments per quarter

The data science team measured a 6.8× increase in churn experiments shipped per quarter, driven mostly by removing the access-ticket bottleneck.

Models that survive contact with production

Final-stage retrain on real data, plus the sealed evaluation harness, kept production AUC within 0.3 of every synthetic-trained candidate — a tighter gap than the team's previous workflow.

Marketing got their week back

Campaign design and creative testing now starts on synthetic data on day one. The agency-side back-and-forth on data access has effectively disappeared.

A cleaner regulator conversation

When ANACOM (Portugal's telecom regulator) asked about Pulse's model-development practices, the team handed over the signed receipts and the conversation was over in one call.

What's running

Doppelset LabBigQuery · dbtVertex AIGoogle Workspace SSOGCP europe-west1

Next story

Lattice Retail — retail & e-commerce

Read →
Try it now

Ship faster. Stop arguing with legal.

Generate your first 100,000 synthetic rows in the next ten minutes. No credit card.