dotData

Here is a fact that data scientists know and almost nobody else does: the algorithm is the easy part. You can download a gradient-boosting model for free, tune it in an afternoon, and it will be roughly as good as the one the company down the street is using. What actually decides whether a model works is the features - the columns you feed it. The signal. And building good features out of raw, messy, multi-table enterprise data is slow, repetitive, judgment-heavy work that can eat something like 80% of a project's time. It is the chore at the center of the profession, and for a long time there was no way around it except to hire more people and give them more coffee.

dotData's entire thesis is that this chore is automatable. The company was established in 2018 as a spinout of NEC Corporation, the century-old Japanese electronics and IT giant, and set up shop in San Mateo, California. Its founder, Ryohei Fujimaki, has a resume that reads like a setup to a joke about overqualification: he was the youngest research fellow in NEC's 119-year history, a distinction held by six people out of more than a thousand researchers. He has a PhD from the University of Tokyo in machine learning. He left all of that to build software that automates the least glamorous 80% of his own former job.

The pitch, which dotData describes as "AutoML 2.0," is a small act of category theft. The first wave of automated machine learning - AutoML 1.0 - automated model selection and hyperparameter tuning, which is to say it automated the easy part. dotData's argument is that the hard, valuable, still-manual part was feature engineering, and that automating it is where the real leverage lives. So the software goes and does it: it crawls across dozens of related tables, including time-series data, and automatically generates and ranks features - thousands of them - looking for the signals a human might take weeks to find, if they found them at all.

The clearest illustration of what that means in practice comes from SMBC, the giant Sumitomo Mitsui Banking Corporation. In dotData's telling, the bank scaled its data science output by something like 40x and ended up generating on the order of two million features a year. Two million is one of those numbers that stops meaning anything at human scale, which is precisely the point. No team of analysts hand-crafts two million features. The only way you get to two million is if a machine is doing the crafting, and a bank is comfortable enough with the results to let it.

That comfort is not incidental - it is a product decision. dotData leans hard on explainability, because its customers live in places like lending, insurance, and financial services, where "the AI said so" is not a legally acceptable answer. Every signal the software surfaces is meant to be traceable and readable in plain language. There is also a quieter piece of engineering discipline underneath the time-series work: avoiding data leakage, the classic trap where information about the future accidentally sneaks into the training data, producing a model that looks brilliant in testing and falls apart in production. dotData's time-series feature generation is built to design around that. It is the sort of unglamorous safeguard that separates a demo from a deployment, and it does not fit on a billboard.

The most interesting recent turn is dotData Insight, launched in 2023, which inverts the usual AI value proposition. Most tools promise to give you the answer. Insight is designed to give you the question. It pairs the company's automated signal-discovery engine with generative AI: the discovery engine finds the statistical signals hiding in your data, and the language model translates those signals into readable business hypotheses about what might actually be driving your KPIs. Instead of a black-box prediction you have to trust, you get a testable idea you can argue about in a meeting. For an executive who has been burned by confident, unexplainable dashboards, "here is a hypothesis worth investigating" is a genuinely different sales pitch.

The money behind all this tells its own story. dotData has raised $74.6 million total, and its 2022 Series B - $31.6 million - was led entirely by Japanese enterprises: Otsuka Corporation and two Sumitomo Mitsui institutions. This is patient, strategic, enterprise money, the kind that bets on infrastructure and renewals rather than viral growth. It fits the company. dotData is not a consumer app; it is plumbing for the modern data stack, and it has spent the last couple of years making sure that plumbing connects to everything - native Snowflake integration, Salesforce connectors, and Amazon Bedrock support that pipes state-of-the-art language models, including Claude, into feature discovery. By early 2026, Insight 2.1 was shipping with support for the newest models on the market.

What makes dotData worth paying attention to is not that it is the loudest AI company - it is nearly the opposite. It is a roughly 88-person, business-to-business software company solving a problem most people outside data science do not know exists. But that problem sits underneath an enormous amount of the work enterprises are trying to do with their data, and dotData has been quietly grinding on it since before "generative AI" was a phrase anyone used at dinner. In an industry addicted to the demo, there is something clarifying about a company that picked the boring 80% and decided to own it.