Deeptech/Artificial Intelligence/Analysis/

The big idea: AI that can “taste” coffee

Israeli-Colombian startup Demetria is developing an AI that can analyse unroasted coffee beans to see what kind of brew they will make.

By Maija Palmer

Welcome to The Big Idea, where each week we pick a company with a weird, wacky or gamechanging notion that could have a massive impact on the world around us if they can turn it into reality.

Next up is an AI that can “taste” coffee, potentially making your morning brew better and solving a big problem for coffee producers.

What’s the problem?

Getting the perfect coffee from the farm all the way to your cup is a multistep process that involves several rounds of pulping, milling, fermenting and drying.

But it’s only towards the end stage of the process that you get any idea of what the coffee will actually taste like.

This final grading of the coffee — known as “cupping” — is the make-or-break moment for the coffee farmer or merchant when a small batch of coffee is roasted and made up into a brew. Only then do you really know if this batch of beans is going to give you the taste profile you want.

This is becoming more and more of an issue in the industry as customer’s tastes have become more exacting, says Sebastián Pinzón. He’s general manager at Carcafe, the Colombian coffee division of Volcafe/ED&F Man — one of the world’s oldest and largest agricultural commodity traders.

“The coffee industry is starting to be very specific about what they want to be in the flavours,” he says. “Coffee is sold as a commodity, but for coffee connoisseurs it is becoming anything but. It is like wine in the variety of flavours.”

This is tricky to achieve, Pinzón explains, in part because a single consignment of coffee can come from 10 to 20 different growers. A coffee exporter like Carcafe has to check that each grower is providing the right profile and quality of beans. It’s also expensive and analogue — Carcafe has a team of 23 cuppers whose job it is to taste coffee all day.

“Cupping is so expensive and exclusive that in general in the coffee industry you just don’t cup enough,” says Felipe Ayerbe, founder and CEO of agritech company Demetria. “It is one of the biggest bottlenecks to getting quality coffee.”

What’s the big idea?

Seeing this problem, Ayerbe had the idea of using AI to break through this bottleneck, developing a set of sensors and a machine learning algorithm that would be able to sample coffee when it was still in its green state to find out how it would taste.

AI is already being used to taste the yeast used in brewing beer and even to create new whisky blends, but it could be an even more powerful shakeup of the $450bn coffee industry, which still largely relies on ancient production methods.

“We want to empower the vast majority of growers who haven’t tasted their own coffee to know more about what they are selling. They can understand the quality of what they have grown and find the best buyers for it,” says Ayerbe.

How does it work?

The sensor technology that Demetria uses has been around for four decades but only in the last five years has it become small and cheap enough to use on a mass scale. The company uses portable near-infrared sensors to analyse the chemical composition of the green coffee bean. The tricky bit is teaching the machine learning algorithm to map that chemical ‘fingerprint’ to a certain kind of taste once the coffee is roasted and brewed.

Ayerbe said the team worked with professional cuppers to calibrate the system, going through thousands of coffee samples. It is painstaking and subtle work, he says. The system had to be taught, for example, to recognise the difference between a molasses sweetness and a chocolate or caramel sweetness in the coffee.

“The model is still learning,” says Ayerbe. “One of the most difficult things is that cuppers will have a lot of nuance in their tasting and there will be a range of coffees that will fit into the taste profile they are looking for.”

It took four months just to create a database of sample profiles.

Demetria coffee sensor

How far along is the technology?

It is still the early days of Demetria and Carcafe’s pilot project. They have now reached a stage where the system can start to be used as part of the quality control process. One of the most basic applications is to detect a defective batch of coffee — that’s already possible, and has value for Carcafe, says Pinzón.

The next stage is starting to teach more of Carcafe’s employees and growers to use the system. Although there was some initial scepticism about the system, Pinzón says cuppers are now starting to turn to the app to adjudicate when they have disagreements among themselves.

Pinzón is clear — the technology will never come in to replace cuppers who will remain the elite quality controllers of the system. But Demetria would help to spread quality control further upstream and to farms and regional hubs where it was never possible to do this before.

What are the risks?

With the technology still in development it isn’t clear how useful it will turn out to be for Carcafe. Pinzón says there could potentially be a return on investment within a year or two if the technology works as they expect it to. But, he says, Carcafe started the project not knowing exactly when it would be completed.

“Noone had done this before, so we really didn’t know how it would go. But it wasn’t prohibitively expensive to do this — it is affordable enough to be worth taking the risk,” he says.

Who is the team behind the idea?

Demetria is a startup based in Israel and Colombia, cofounded by Felipe Ayerbe, a former investment banker and consultant, whose grandfather was a coffee farmer who brought his harvest to market by mule.

His cofounders include Eduardo Shoval, an Argentine-Israeli serial high-tech entrepreneur and angel investor; Salomon Kassin, a commodity trading expert and pioneer of Colombia’s speciality coffee industry; and serial high tech entrepreneur Yori Nelken, a pioneer in natural language processing.

Demetria raised a $3m seed funding round in March.

Maija Palmer is Sifted’s innovation editor. She covers deeptech and corporate innovation, and tweets from @maijapalmer

Join the conversation

avatar
  Subscribe  
Notify of