I’ve spent time with data science teams who built models they were genuinely proud of. The experimentation was thoughtful. The validation was solid. In development, the models performed well. But when it came time to move from notebook to production, things slowed down. Not because the model was flawed — but because the path around it wasn’t clear.
What I’ve come to notice is that many organizations don’t have a modeling problem. They have an engineering and alignment problem. A model that works in isolation still needs integration into applications, production-grade data pipelines, monitoring, scaling, retraining workflows, and ownership. That work usually falls to engineering teams who are already committed to other priorities. So the model waits. Sometimes for weeks. Sometimes indefinitely. And the cost isn’t just delay — it’s the gap between potential value and realized value.
I don’t think this is about blame. Data scientists are often measured on model performance. Engineers are measured on delivery and reliability. Product teams are measured on shipped features. Each group is doing what they’re incentivized to do. But unless there’s shared planning around how a model becomes a maintained capability — not just an experiment — the work fragments. The model may be “done,” but the system around it isn’t ready.
The teams I’ve seen make progress tend to think about production earlier. They define deployment patterns before scaling experimentation. They clarify ownership across data science, engineering, and product. They invest — sometimes uncomfortably — in reusable pipelines and automation that make the next deployment easier than the last. It’s not glamorous work, and it doesn’t always show immediate ROI. But over time, it reduces friction between functions.
I’m still learning where the right balance is. There’s always pressure to ship the current model quickly. But I’m starting to believe that the real leverage isn’t in building more models — it’s in making it normal for them to leave the notebook.
Leave a Reply