Ask a room full of ML engineers what’s hardest about their job and you’ll get a hundred answers about model architectures and hyperparameter tuning. Ask the people who run the teams, and the conversation shifts immediately to something far less glamorous: infrastructure, data plumbing, and the unglamorous work of making models survive contact with production.
Omar Rahman, who leads ML and data engineering for defensive security at Salesforce, is blunt about where the friction lives. Productizing ML, he says, requires far more than training a model. Serving predictions, detecting when data statistics drift, and adapting to those changes creates what he calls significant overhead. His other pain point will sound familiar to anyone who’s worked inside a large enterprise: data is siloed and poorly maintained, which means teams burn enormous amounts of time on consolidation, preprocessing, and cleaning before any actual ML work begins.
Danica Xiao, Senior Director and Head of Data Science and ML at Amplitude, brings a researcher’s pedigree to the conversation with more than 95 published papers. Her team builds self-service ML products on top of multi-sourced user data for digital product analytics, which puts her at the intersection of two hard problems: making models that work across messy, heterogeneous data and making them accessible to non-experts.
Rahman’s most pointed observation may be his diagnosis of the most common mistake he sees: teams reaching for the newest algorithms and freshly open-sourced libraries while ignoring the business problem in front of them. A lot can be achieved, he notes, with simple traditional ML techniques. That’s a message worth amplifying in an industry where the release cycle for foundation models seems to outpace most teams’ ability to even evaluate them, let alone deploy them responsibly. He sees the real payoff of generalized AI capabilities in areas like healthcare and transportation, where radiology is already seeing meaningful manpower savings and self-driving vehicles are reshaping logistics. The throughline from both directors: the hard part was never the math. It’s everything around it.
