For three years I watched it happen from a consulting seat. Clients who used to ask about churn models and demand forecasts started asking about RAG pipelines and chatbots. The energy left the room whenever tabular data came up. Everyone wanted to do something with their PDFs. Nobody wanted to touch the datasets where the actual business lives: the millions of rows and columns of tabular data that underpins every enterprise.
I understood the pull. Working with LLMs feels like magic when it comes to text, images and code. In contrast, tabular ML is a hard sell: months of feature engineering, hyperparameter sweeps, brittle pipelines, and at the end of it a model that maybe moves a KPI if you're lucky and your data is clean (which it usually isn't).
I believed in tabular ML because I'd seen it work first-hand. My first job was at Iwoca, lending to small businesses. The credit scoring and loan sizing models my team built added millions in monthly revenue and cut defaults at the same time. Structured data. Tabular features. No LLMs in sight. That experience set my baseline for what predictive analytics can do.
But I soon realised that Iwoca was the exception, not the rule: most companies have the data, have the use case, and still can't get a model into production that anyone trusts. So they give up and start building chatbots instead.
I believe things are about to change, however thanks to the advent of Large Tabular Models.
Fundamental built one. It's called NEXUS. It does to tabular ML what GPT did to NLP: collapses weeks of bespoke work into a single pretrained model that works on your structured data out-of-the-box. No feature engineering required. No hyperparameter tuning. And unlike the XGBoost you tuned in 2022, it gets better every time the team ships a new version. Performance gains compound at a pace that would have been unthinkable just a few years ago.
When Fundamental asked me to join as a Forward Deployed Engineer, I said yes the same week.
So is the FDE role about to disappear too? If the model handles itself, what's left for us?
A lot, actually. It’s just different work.
The model doesn't know what problem you're solving. Translating a vague business ask ("we want to reduce churn") into a crisp ML problem with a real success metric is still a human job, and most people are bad at it.
The model doesn't know your data either. Even with foundation models doing the heavy lifting, how you prepare and structure inputs is still where the last meaningful gains live.
Finally, NEXUS works really well out of the box. That's the research team's job, and they're relentless about it. Our job as FDEs is to take that baseline and push it further: refining the model for every client so it performs at its peak on their data, in their environment, in service of solving their specific problems.
The shape of the job has changed. The leverage has gone up. One person can now do what a team of five used to do (usually quite badly).
We're hiring FDEs at Fundamental. The job is sitting between customer problems and a model that's genuinely better than what came before, and turning that into production systems that move real numbers. If you've felt the same frustration I have, watching tabular ML get abandoned for the shiny thing while knowing how much value it can actually deliver, come build with us.
