6 production models, deployed by non-technical teams.
A Fortune 500 travel company had the data to personalize every trip it sold, but only a small data science team could turn that data into models. Product managers queued for weeks, marketing ran on broad segments, and the recommendation engine lagged on conversion. Plexe gave product and analytics teams a way to build and ship production models themselves, and replaced the incumbent recommender with one that won its A/B test.
The challenge
- PMs waited weeks for data science to build models
- The recommendation engine underperformed on conversion
- Marketing ran on broad segments, not ML
- Insights were locked behind technical skillsets
What we built
- PMs and analysts deploy models independently
- A new recommendations model, live and A/B tested against the incumbent
- 6 production models powering targeted campaigns
- A self-serve platform for non-technical users
Results
- +7% conversion uplift, A/B tested
- 6 models live, run by PMs and analysts
Why they came to us
At Fortune 500 scale, small changes in conversion are worth a lot, and the company knew its recommendation engine was leaving money on the table. The bigger problem sat upstream: every model request, from a new ranking to a campaign audience, went through one central data science team. Product managers waited weeks for work that should have taken days, and most ideas never made it into the queue at all.
Marketing felt it most. Without models of their own, campaigns ran on broad, rule-based segments. The insight existed in the data, but only people with a specific technical skillset could get it out.
Goals and success criteria
We agreed three outcomes up front. First, beat the incumbent recommender on conversion in a controlled A/B test, not an offline benchmark. Second, let PMs and analysts build, deploy and own production models without filing a ticket with data science. Third, get those models into live marketing campaigns, so the work paid back in revenue rather than in slide decks.
How it works
The centrepiece is a new recommendations model, trained on the company's own booking and behaviour data and deployed alongside the existing engine. It ran head-to-head against the incumbent in a live A/B test and delivered a 7% uplift in conversion before it was promoted to production.
Around it, Plexe set up a self-serve modelling workflow for non-technical teams. PMs and analysts describe the problem, pick the data and ship the model; data science reviews where it matters instead of building everything by hand. Six production models now run this way, powering targeted campaigns that used to rely on static segments.
What they own now
The models, the pipelines and the workflow all live inside the company's own environment and are run by its own teams. Data science moved from being a bottleneck to being a reviewer, and the backlog of model ideas is now something product and marketing can work through themselves.