Engineered intelligence

Outthink Human Intelligence.

Now with VEOX you can invent algorithms that otherwise humans can't — continuously engineering better intelligence from your data.

The ceiling nobody measures.

Every organisation knows what its models score. Almost none know what they could have scored. That number has never been anyone's job.

Deadlines, not limits

Your model stopped improving on a Tuesday

Someone shipped it because the quarter was ending. That date was a calendar decision, not a mathematical one — and the model has sat at that number ever since.

The gap is invisible

There's no line item for accuracy you didn't reach

It doesn't appear in a budget. It appears as claims leakage, unplanned downtime, churn you priced as normal, and inventory you wrote off as the cost of doing business.

But it's searchable

Given a score and enough compute, the space above your baseline can be explored

Not guessed at, not benchmarked against a vendor's demo — searched, measured, and handed back to you as something that runs. That's the entire company.

The category

Engineered intelligence.

There's a name for what happens when you stop hand-building models and start searching for them. We've been calling it engineered intelligence: a system that proposes candidate algorithms, scores them against your data, keeps what wins, and doesn't get tired.

The distinction that matters is where the intelligence sits. In every previous generation of tooling, a person decided what to try and the machine executed it faster. Here, the machine decides what to try. A person sets the objective and the budget.

It's a young category and we're not going to pretend otherwise. What we can do is publish the record, including the runs we lose, and let you check it.

Frequently confused with
What actually happens
AutoML — fits the model families you named, faster
Searches structures nobody named, including ones that don't have a name yet
Fine-tuning — adapts a fixed architecture to your data
The architecture is the thing being searched
A dashboard — shows you what already happened
Returns runnable code you own and deploy
An agent — calls tools on your behalf
Produces artefacts that outlive the session

One engine, several outputs

jAIn is the first thing we've shipped. It won't be the last.

Everything below runs on the same search. What changes is the space it searches and what it hands back.

The record so far

Published, including the losses.

A leaderboard that only shows wins isn't a leaderboard. Every run we've scored is in the table on the jAIn page, wins and losses in the same list.

847
real datasets scored against the model a team had already built
71%
of runs where the search found something better than the incumbent
+2.4pt
median improvement on the runs it wins, on the client's own metric
29%
of runs where the hand-built model held. We publish those too
placeholder figures — swap before launch See the full table →

VEOX Research

Read the argument before you take the call.

Both papers are open, ungated, and written for someone who'll check the method. Start with the technical one.

Who's building this

Small on purpose.

VEOX was founded by Jepson Taylor — ex-quant, ex-HireVue, holder of 14 AI patents. He sold his deep-learning company Zeff.ai to DataRobot, then held C-level roles at DataRobot and Dataiku. He has watched two generations of AutoML get sold and knows exactly what this isn't.

The engineers behind him came from Intel, Micron, Cisco, VMware and Adobe. We work with a deliberately small number of design partners at a time, and we're selective about fit — if your problem is one the search won't help with, we'd rather say so on the first call.

IntelMicronCisco VMwareAdobe DataRobotDataiku

Find out what your best model left behind.Start with one table.

Evaluating VEOX for a team? There’s a page for your security review →