AI on OPT: Design Data Products

The last article was important to start the engine, and I'm grateful for the support I'm receiving. Now, we definitely should explore more about how we build AI products. The question is: how do we go from an exciting idea to a fully functional AI product? Whether you’re aiming to predict player performance in NFL or optimize business operations, the steps to build an effective AI solution are surprisingly similar. In this article, I’ll walk you through the key stages of designing AI and data products, helping you understand how to go from idea to implementation. And since I really like football (soccer for my American friends), I’ll use an example from the Italian league to explain how this process works.

Define Objective

The first step in any AI project is to clearly define what you want to achieve. This could be anything from forecasting demand for a product to predicting how well a soccer player will perform in the upcoming season. Having a well-defined objective not only sets the direction for your project but also helps you decide on the right tools and methods to use.

A Serie A club’s data analytics team might set an objective to scout the best players across leagues, predicting which players are most likely to succeed in their team based on performance metrics, physical attributes, and even behavioral data. Pretty much like Moneyball.

PS: If you're on OPT and reading AI on OPT, come close, and let's hug and be friends. I feel you. And if you have a job already - good job, soldier - remember your fellow OPTs and reach out when you see an opportunity in your company. No AI can replace humans on this matter (I... think?).

While writing this article, I listened to Lucio Dalla.

Enjoy my fav song.

I'm Andrea, a Senior Product Manager with 5+ years of experience building B2B SaaS solutions.

If you liked my article, feel free to share it on your social media.

Ciao!

It's not all about business.

Ok, enough with business for this article. Let’s have a look at a few cool AI-related videos I found:

• Building an Advanced AI Basketball Referee (Ayush Pai)

• A.I. ‐ Humanity's Final Invention? (Kurzgesagt – In a Nutshell)

• AI Art: How artists are using and confronting machine learning (MoMa)

Designing AI products is an exciting journey, with each step bringing you closer to impactful solutions. By understanding the process and making smart choices with technology and data, you’ll create effective and engaging AI.

Next time, we’ll discuss traditional machine learning vs. neural networks and (maybe) explore the world of generative AI. Stay tuned!

Andrea

PS: I wrote most of the content in this article by taking inspiration from Lutz Finger's classes I took during my MBA. I want to express a sincere appreciation for what he taught us. He just released a great course on eCornell. Check it out!

"We must be sure we’re working with valuable data.

The equation is simple: sh*t in = sh*t out"

Collect & Clean Data

Once you have your objective, the next step is to gather the data you’ll need to reach it. Data can come from various sources, like historical records, live feeds, or external datasets. However, raw data is rarely ready to use straight away. You’ll need to clean it—removing errors, filling in gaps, and making sure it’s consistent.

The scouting team might collect data from various leagues, including player statistics, match outcomes, physical fitness records, and scouting reports. Cleaning this data could involve standardizing formats, correcting any inconsistencies, and dealing with missing values.

Develop and Train & Test Model

With clean data in hand, it's time to develop your AI model. This is the core of your AI product, where the data is used to make predictions or decisions. Developing a model involves choosing the right algorithms, training it on your data, and then testing it to see how well it performs.

Here’s where the choice of technology comes into play. The decision on what technology to use—whether traditional machine learning or more advanced neural networks—depends on several factors:

• Objective: What are you trying to achieve? Simpler tasks might be well-suited to traditional machine learning, while more complex problems might require the power of neural networks.

• Type of Data: Do you have structured data, like numbers in a spreadsheet, or unstructured data, like images or text? The nature of your data often dictates which approach will work best.

• Interpretability: How important is it for you to understand how the model is making its decisions? Traditional machine learning models often offer more visibility (interpretability), allowing you to see exactly which factors are influencing the predictions. In contrast, neural networks can act like a "black box," providing powerful predictions but with less transparency about how they arrived at those results.

If the scouting team wants a straightforward model to evaluate players based on clear, structured stats, traditional machine learning might be the way to go. However, if they’re working with complex data like video footage from matches and want the model to learn from patterns in the videos, they might opt for a neural network, even if it means sacrificing some interpretability for accuracy.

UX & User Testing

Creating a powerful AI model is one thing, but making sure it’s usable and valuable to end-users is another. This is where UX (user experience) design and user testing come into play. The goal is to present the AI’s insights in a way that’s intuitive and actionable, whether through dashboards, reports, or even automated suggestions.

The scouting team might design a user-friendly dashboard that presents the model’s predictions, allowing coaches and managers to quickly identify which players are the best fit for their team and make informed decisions accordingly.

Productionalize / Visualize / Feedback Loop

The final step is to deploy your AI model so it can be used in real-world scenarios. But deployment isn’t the end of the road. You’ll also need to visualize the results in a clear, understandable way and establish a feedback loop to continuously improve the model over time.

After deploying the player scouting model, the team might integrate it into their decision-making processes, using it to inform player acquisitions and match strategy throughout the season. Regular updates and new data would feed back into the model, keeping it accurate and relevant.