PyMCJAXNumPyroDatabricks

The Challenge

Haleon already had a capable internal team building probabilistic models for marketing analytics. They didn't need someone to build models for them — they needed a sparring partner. The team had hit several technically thorny problems: handling missing data in models with complex hierarchical structure, getting their code to run efficiently on GPUs in a cloud environment, and making sound architectural decisions about time series models. These are the kinds of problems where having someone who's solved them before saves months of trial and error.

Our Approach

We worked with Haleon's team in a coaching format — biweekly calls structured around code review, architectural questions, and working through specific technical roadblocks. When the team ran into trouble with missing observations in their hierarchical models (a genuinely subtle problem), we walked through the options and tradeoffs rather than just handing over a solution. We helped them get GPU-accelerated sampling working in their cloud environment and provided architectural guidance on their time series modeling approach. The goal throughout was to make ourselves unnecessary.

Results

By the end of the engagement, Haleon's team was self-sufficient — building, validating, and deploying their own models with confidence. The best measure of success in a coaching engagement is when you run out of meaningful critiques, and that's exactly where things landed.

The relationship also opened doors to other analytics teams within Haleon, as we presented our approach to their broader organization.

“PyMC Labs significantly enhanced our testing capabilities by leveraging Bayesian programming, with invaluable advisory support driving substantial operational improvements. Their advisory role in delivering new feature requests and training has been invaluable.”

Nathan Kafi , Principal Data Scientist, Haleon

PyMC Labs Team

  • Juan Orduz
  • Oriol
  • Niall Oulton
  • Christian
  • Adrian

Let's Chat, We Respond Fast

Tell us about your problem. We typically respond within 24 hours.

Schedule a Consultation