)
VIDEO: Automating the Hard Parts of Data Science
I sat down with Razi Raziuddin, co-founder and CEO of FeatureByte, and we kept coming back to the same point: building the model is not the problem. The real drag is everything before it: understanding the data, engineering features, building pipelines that hold up in production. That's 90 percent of the work, and it's the part that has stayed largely manual.
Razi came up through DataRobot, same as me. When he and co-founder Xavier Conort (former DataRobot chief data scientist, ex-number-one on Kaggle) left to start something new, they went after feature engineering specifically because it was the biggest unsolved problem in the lifecycle.
What FeatureByte Does with Feature Engineering
FeatureByte is a data science agent. You describe the use case, point it at your data warehouse, provide a target variable, and it takes over: analyzing metadata, generating and evaluating features, handling point-in-time correctness to avoid target leakage, and building models. Two to three months of work gets down to two to three days, with model performance improvements of 3 to 18 percent from better features alone.
Feature Stores and the Hype Gap in Data Tooling
Feature stores are a useful example of the gap between tools people talk about and tools teams actually use. The concept is sound: a centralized, reusable library of engineered features. In practice, Razi noted that real-time use cases like fraud detection genuinely benefit from them, but for the other 95 percent of workflows, a well-organized table does the same job.
The pattern is consistent across data science tooling. Razi compared it to the Hadoop era, when every project spawned a new tool with an animal name. A few became foundational. Most disappeared. The ones that survived solved a real, recurring problem in a way teams could actually adopt.
Why Data Science Automation Is Not Optional
Razi said a lot of the data scientists he talks to are trying to figure out whether to pivot toward AI engineering or keep doing what they have been doing for the last decade. Meanwhile, the business side wants more models in production, updated more frequently. Without automation, nobody is keeping up with that.
I ran a live coding webinar where a co-founder one-shotted a full EV charging station optimization problem in a single prompt and got everything needed to run it. The output quality from these models has advanced significantly, and teams that are not building automation into their workflows are already falling behind.
Where Tabular Data Fits in an LLM-Driven World
Razi's take on how better foundation models affect FeatureByte was direct: tabular data is a different problem than language. The same number means something completely different depending on the column, the table, the business, and the industry. There is no inherent shared context to train on. Tabular foundation models are still a research-stage concept, and classical ML algorithms continue to outperform them in production.
What does work is a layer of predictive models that translates historical data into signals an agent can use. Pointing an LLM at three years of transactional history and asking it to reason over that data does not work. The agent framework that sits around well-built predictive models is where the real leverage is.
Frequently Asked Questions
What is feature engineering in machine learning?
Feature engineering is the process of transforming raw data into inputs that a machine learning model can actually learn from. It involves selecting relevant variables, creating new ones from existing data, and structuring historical data in a way that reflects real-world timing. Most data scientists consider it the most time-consuming part of building a production model, often accounting for 80 to 90 percent of total project time.
How long does it take to build a production-ready machine learning model?
For most data science teams working manually, building a production-ready ML model takes two to three months from raw data to deployment. That timeline includes data exploration, feature engineering, model training, evaluation, and pipeline setup. Automated data science platforms that handle feature generation and model building can compress that to two to three days.
What is a data science agent and how does it work?
A data science agent is software that automates the end-to-end machine learning workflow without requiring a data scientist to write code at every step. You describe the prediction problem, connect the agent to your data warehouse, and it handles feature ideation, statistical evaluation, dataset construction, model training, and pipeline deployment. The agent uses a combination of semantic understanding and large language models to generate features relevant to the use case, then validates them against the actual data.
Do feature stores actually improve model performance?
Feature stores are most useful in real-time serving scenarios, like fraud detection, where features need to be computed and delivered at low latency. For batch-oriented use cases, which represent the majority of production ML workloads, a well-structured database table accomplishes the same goal. Teams with large data science organizations benefit from feature stores as a way to share and reuse engineered variables across projects, but adoption outside of financial services and fraud use cases remains limited.
Can large language models replace predictive machine learning models for tabular data?
Not reliably. LLMs are not designed to reason over large volumes of historical tabular data, and they lack the ability to perform the statistical computations that predictive models require. The same number carries a completely different meaning depending on the column, table, and business context it comes from, which makes it difficult to train a general model on tabular data the way you can with language. Classical ML algorithms still outperform LLMs on tabular prediction tasks in production. The more effective approach is using predictive models as a translation layer between historical data and the LLM-based agents that consume it.
