From the ground up. Creating AI sportstech for grassroots teams.

Published in

Mingle Sport

8 min readApr 6, 2022

Ball sports are superior compared to any other form of sport. Period 😇. It’s playful, It requires team collaboration. There is healthy competition. And it’s engaging to watch. There might be cultural preferences which ball sport is the most popular, but in general ball sports are on top of the popularity lists.

But grassroots ball sports are amongst the least captured sports. Whereas fitness, running, cycling and many of your social moments are captured and tracked using digital devices, for grassroots ball sports this is not the case.

Practical problems are in the way. How do we track player performance in an easy and affordable way? How do we capture and share video while respecting the privacy of those on the pitch? What business model supports the sports ecosystem that depends so much on volunteers and local clubs?

We started a company called Mingle Sport that wants to tackle these big challenges using ubiquitous mobile technology, AI and the power of social networks. Like many startups our ambitions are huge, but we aren’t naive about the long and painful journey that is ahead.

Note that this post is written at the moment we are in our alpha testing phase. We have spend over a year on R&D and we’re months away from shipping to the App Stores. It’s a hectic time full of insights, rapid iterations and bug fixes.

The principals on which we build our product

Our mission is to let more people enjoy ball games. Doing it more, watching more grassroots moments that matter. Getting better.

Team sport participation is challenged and in many countries on a steady decline. Why? Because individual exercises have fewer barriers to entry. Because gamification and social technology is designed around screens and individual performance. We believe that the forces of individualisation, digitisation and urbanisation can also be used in favour of team sports. If we build products on the following principles:

1. Use what is already there. Unlock premium sportstech using mobile devices people already own: phones, wearables, sports cameras
2. Privacy by Design. Users should be in control of what they share with whom.
3. A fair play business model. A business model that benefits the whole ecosystem in a fair way
4. Build from the ground up: start solving simple problems with tools people are familiar with. Innovate step by step from there.

Pitching the Mingle Sport App

We want to be the default app for ballsports. If your match or training is not on Mingle, it didn’t happen. So, here is our MVP pitch for you.

Our Mingle Sport product teaser

The first version of our product is from a user perspective quite basic. The setup and tech stack behind it though is quite sophisticated. Why and what about that in a minute..
To summarise our MVP let’s you:

Capture Matches (by adding match facts, text, photo, audio and video)
Chat with your team
Profile & Stats

We will start with a club based roll out mechanism. Meaning the app is available to clubs we have ‘unlocked’ on the back end. Not because our app is designed around clubs necessarily. It works for any team or even individual player. But more because we want our app to work on a micro network level. From a user perspective, but also from a tech perspective: our app is all about applying computervision AI to video in near real time and we don’t want to screw that up for the user.

We start with football and are in the process of adding padel, tennis and basketball. In case you want to be an early user, please join the waitlist and we get in touch.

Under the hood of our fancy sportstech product video

Okay, that video looks pretty decent. But lots of familiar features. So what is under the hood?

A team of 26 sportstech professionals

That is a pretty sizeable team. I don’t know that many early phase sportstech startups of this size actually. Especially not in Europe. Let me break down our team setup for you. We have a native Android and a native IOS team. That’s 2x3 people. We have a back end & web team of 4. We have a Machine Learning / Data Science team of 4. Our UX & User research team is 5 + 1 video & motion artist. Our marketing team: 2. Than we have the 5 person founding team that is all over the technical, design, product and marketing place. Finally we do have some great interns plus a few open positions.

The MVP could probably have been build by a team less than half the size. But what you see is not what is coming.

Building an ML Sports Tracker that works with mobile phones

Whereas elite ball sports are almost over-equipped when it comes to sportstech, at a grassroots level things are messy. No stadium, no wired internet, no fixed camera positions. No expensive wearable sensors nor a dedicated sport data analyst. No nothing.

Based on the principle that we use what is available, our vision centers around computervision AI using mobile phones or popular action cameras as input devices. Keep in mind, that there are many very sophisticated cameras available for grassroots ball sports, the problem is that people keep them in their pockets.
You might wonder why we are betting hard on using cameras as a primary tracker opposed to wearable sensors. First, we do not not believe in sensors. They’re great, but the problem with sensors for ball sports is: they do not track one of the most important objects: the ball. Second, in many ball sports you are not allowed to use the smartwatch you already own and so you are forced into using a weird and expensive sensor just to track the ball sport activity. So we will support Apple and Google watch OS and integrate with 3rd party sensors, but our own track development will focus on computer vision AI.

Ultimately our AI is able to track the most important things that happen on a pitch using your smartphone. The only thing we cannot measure is your heart-rate.

Where do we currently stand?

Ball detection for tennis and football
Player tracking for tennis and football.
Sport object detection (goal, cornerflag, etc)
Court detection for tennis

The big milestones we are working on are event detection and player re-identification. Especially the latter will be a breakthrough for sports like football and basketball as there are so many collusions between players and players going in and out of the camera positions. Additional complexity for us is the fact we need to rely on user generated footage from different people.

So, the road ahead is still long. However we already have developed some nice use cases. Let’s look at two practical AI capabilities we currently support:

Player and ball tracking for tennis

Because tennis is a very structured sport we can quite accurately calculate player and ball position using the input from a single mobile camera.

Adaptive Zoom for football

Especially in football the problem with mobile footage is that is shot from too far away. Because our AI knows where the balls, players and goals are we are able to dynamically zoom in and out. It makes the footage much more engaging to watch.

The science behind our adaptive zoom feature

These are just a few examples. Basically our ML team works on two tracks: a core R&D track (the ‘slow track’) where we improve our core models. And the second — product — track where we create practice features that can be used in our app. The better the core, the more use cases and features we can support.

A decentralised setup

Although our ML Sportstracker is one of our key assets it’s definitely not the only one. I’d say that the way we are designing our technology and team architecture is also a key enabler for success.

Feature Teams

26 people. That’s an early phase startup with the size of a scale up. It’s too big for a single team, so after quite a few iterations we landed on a feature team setup. A feature team has all the key roles needed to deliver a feature. UX, back end, Android, IOS, web and product owner. Currently we have 3 feature teams:

Match & Activity
Profile and Stats
Team chat & collaboration

As the Computervision / ML team works mostly on the R&D track they are currently posisitioned as a seperate product team. However they do get pulled into production feature teams every now and then. Some people join multiple teams, but for the most part these teams are small enough for everyone to make an impact.

Microservices and tech stack

Our front end is a native IOS and Android app plus a web app. On the back end we support this with a variety of microservices. The most important ones are our user service, match and club services, our video service and messaging service. We leverage various clouds (AWS, Azure and in the future Google as well) for these services.

User Research

Although we love to talk about our product of course we understand we first and foremost need to solve real user problems. Here is our arsenal of user research tools we apply:

Focus Groups & Interviews: exploring user needs and early testing of key hypothesis. We do this remote and on site
Competitive research: testing apps and tools from other sportstech players in order to learn from them
Prototype tests: testing app prototypes, video formats, camera setups and so on.
Surveys: for quantitative research
UX tests: doing quick UX tests to test specific flows

Although all of these tests combined provide us with a pretty clear view of the key direction, the real tests need to happen on production where things actually come together in a real life setting. And since we are building apps with sophisticated social and ML technology it’s a challenge to build very fast. Our key strategy to mitigate is to focus on micro networks (clubs) first and scale fast when our app really works well at a few clubs.

We are super excited about our journey. Although it’s a very humbling journey and probably with some bumps on the road we are convinced of the opportunity sportstech holds for grassroots ball sports.

If you want to learn more about us you can actually become a ‘Friend of Mingle’ and get access to insider updates. It’s invite only so contact me on LinkedIn.