Why Data and AI Still Break at Scale (and What to Do About It)

​It’s 2025, and most teams still struggle to get data and machine learning projects working across people, systems, and time.

​Notebooks don’t hold up. Pipelines are brittle. Results drift. Sharing work often means breaking it.

​Akshay Agrawal has dealt with this across environments—doing ML at Google Brain, optimization research at Netflix, academic work at Stanford, and now building marimo, an open-source Python notebook reimagined for modern workflows—versionable with Git, executable as code, and easy to share as apps.

​In this episode, he joins Hugo Bowne-Anderson to talk about why data and ML work so often falls apart—and what’s required to make it hold together.

​We cover:

• ​⚠️ Why reproducibility and collaboration still fail on real teams

• ​🧠 What gets lost moving between research and production

• ​🔧 How small tooling choices create large-scale failure modes

• ​🧭 What Akshay has learned building across ML, research, and open source

​If you’ve ever tried to get your team’s work running somewhere else—and watched it fall apart—this episode will feel familiar.