Six slides on the Main Stage

In November 2021 Apres was picked as a Breakout Startup at Web Summit in Lisbon, and I presented it on the Main Stage. The deck file is dated October 25, 2021. It has six slides. Here they are, with what I think of each one now.

1. The title

The Apres mark on an iridescent color field. No words at all, which is right for the first ten seconds.

2. The problem, in three words each

The headline was “AI has a learning problem.” Under it, three icons with three labels: Siloed Expertise. Continuously Degrades. Hard to Understand.

This is the slide I would keep exactly as it was. Anyone in the room who had shipped a model would have recognized all three. The people who know what a model should do sit in a different department from the people who build it. The model gets worse after launch, quietly. And almost nobody can tell you why it made a given decision.

3. The workflow everyone recognized

The left side was the typical machine-learning workflow in boxes: acquire data, explore it, prepare it, engineer features, select and train a model, tune it, evaluate it, predict. Below the predictions sat a grey band with three boxes: compliance documentation, decision explanation and model monitoring. Then the decision goes to the business or the customer, an action is taken, and the loop starts again.

The right side was black, with four news headlines: Amazon scrapping a recruiting tool that showed bias against women; a self-driving Uber car killing a pedestrian in Arizona; JPMorgan agreeing to pay $55 million to settle a mortgage discrimination complaint; Twitter’s image-cropping algorithm appearing to have a white bias.

The grey band was the point. Everything after the prediction was where Apres worked.

4. The loop

“The distributed learning platform,” with one line under it: “Data → Feedback → Intelligence.” Below that was the product, on a customer churn example: an inbox of insights, alerts and suggestions, a plain-language narrative about a model’s “10 Best Customers,” and a box asking the user a question:

Do you confirm that these accounts can be described as your “10 Best Customers”?

The Apres app, on the Web Summit slide, 2021

That box is the part of the product I would still build. The model makes a claim in plain words, and a person who knows the business confirms or corrects it. The correction feeds back into the system.

5. Growing fast

Partner logos, a row headed “Supported by” (Faber, Semapa Next, Techstars, Adara Ventures) and two press callouts: WIRED’s 100 hottest startups in Europe, and TechCrunch naming Apres a top ten startup at Techstars’ remote demo day.

This is the slide I would change. It had logos and quotes about us, and not one number from a customer. Apres shared results in public posts: a 44% improvement in model performance for one of the largest banks in the world, and a 20% average improvement across customers. One line like that would have done more than a row of logos.

6. The close

Logo, website, my email.

What carried over

Looking at slide 4 now, I see the shape of what I work on today. Excellent’s About page has a formula on it: verified = claim + evidence. At Apres the model made a claim and a person who knew the business confirmed it. At Excellent the agent makes the claim, and the evidence comes from the repository, the tests, the build and the browser. The loop is the one I put on that stage in 2021: a machine says something, and something independent checks it before anyone acts on it.

Six slides was the right number. I would just trade the logos for the number.