# Shipping a football analytics app for the 2026 World Cup with agents

A football analytics app went from concept to World Cup launch in under six months and shipped 22 releases, but we had our challenges.

- Author: Dan Mocan (https://www.linkedin.com/in/dan-mocan-9424421a/)
- Published: 2026-11-16
- Kind: Article
- Page: https://x2labs.ai/notes/world-cup-football-analytics-app

**TL;DR**

- The founders had a professional-level football analytics idea and one launch date they could not move: the 2026 FIFA World Cup.
- We built an iOS and Android app powered by live match data: scores, player ratings, deserved win, momentum, field tilt, phase-of-play stats and shareable visual stories.
- We integrated and adapted around 40 machine learning models built by one of the founders, a football and data intelligence expert. The models turn raw match events into stories about players, teams and matches.
- The result: under six months from concept to launch, followed by 22 App Store releases between 8 June and 1 October, with a 4.8 App Store rating (36 ratings, October 2026).

The founders wanted fans to see more of what is actually happening in a match, using the kind of analysis normally available to professional clubs.

Who deserved to win. Who controlled territory. How each team built up, progressed and finished. Real data, directly in the hands of fans.

The idea itself was clear. The problem was timing.

The World Cup gave us a hard deadline. There was no option to launch a few weeks later and treat it as a normal delay. By then, a big part of the opportunity would already be gone.

The second problem only became obvious once we started working with live matches: the data is very fluid.

A goal can appear in the feed and then disappear after VAR. The scorer can change. Events can arrive late or be corrected afterwards. Different providers sometimes represent the same thing differently.

That matters in a consumer app because people notice. They compare what they see with the match itself, with other apps and with screenshots shared online.

So we had two major risks: missing a deadline that was only a few months away, and showing numbers people could not trust.

## What we built, and who decides what

We built a Flutter app for iOS and Android and a backend that ingests a live event feed for every match. And we have built it using our agentic development workflow, with Claude Opus and Claude Fable as the development agents, from February to October 2026. We always used the latest version available: Claude Opus 4.6 from February, then Opus 4.7, 4.8, 5 and 5.5 as each came out, and Claude Fable 5 and 5.1 from June.

Everything fans see sits on top of that data: deserved win and momentum, field tilt, player ratings, performance across more than 10 different phases of play: buildup, progression, finishing, fast break, transition, set piece, etc alongside styles of play. Fans can rank teams and players by stat or phase.

![The match screen for LAFC 2–2 LA Galaxy: match rating, deserved win 58% to 20%, possession, expected goals and shots, above a momentum chart and the expected goals timeline](https://x2labs.ai/_astro/match-overview.D4PrggWs_HinAs.webp)

*The live app with real data from an MLS match, LAFC 2–2 LA Galaxy. The app launched for the World Cup; we added MLS and the Premier League after the tournament.*

![Three match views: a shot map with shots, big chances and dangerous attacks; a territory map with 71% to 29% field tilt; and an action heat map above the phases of play](https://x2labs.ai/_astro/shots-territory-phases.Cc3zOUVD_1yApIB.webp)

![League views for Major League Soccer: the team table with points and ratings, team rankings by overall, attack and defense rating, and players ranked by rating](https://x2labs.ai/_astro/league-rankings.ihOxWc2I_Z1TgbWC.webp)

The product consists of around 40 ML models built by one of the founders. Our job was to integrate them into the live pipeline and adapt them to real match data.

Each model has a narrow role. Some explain how players and teams performed. Others turn the numbers into visual stories fans can share. All of them work from the events of a specific match.

![Team tendencies for LAFC on sliders such as patient versus direct buildup and control versus chaos, next to success rates for buildup, progression, finishing, fast break, transition and set piece phases](https://x2labs.ai/_astro/team-tendencies-phases.CSikiX67_29eREi.webp)

Humans are still in the process in key points like:

- **Before a feature ships.** Someone decides which metrics and stories are useful and how they should appear on a phone screen.
- **Every matchday.** QA compares results, standings and key events in the app with independent public sources.
- **Before every release.** Shipped code is reviewed and validated by our senior developers.

This matters because a silent error can travel quickly. From what I’ve seen, media now quotes deserved win and field tilt metrics in match recaps.

## What didn’t work

**Treating one feed as the absolute source of truth.** We initially assumed that a licensed feed would simply be correct. Most of the time it is, but our World Cup experience taught us how much live data can change. A correct lineup sent 30 minutes before the game start might then change just before kick-off because of an injured player. You have to build for those corrections and accept that some edge cases will only become clear when you encounter them.

After that, we started checking anything that could produce a plausible-looking but incorrect result much more aggressively.

**Expecting a goal to be a clean event.** We assumed a goal would arrive as a minute and a scorer. Often it does not.

Goals get credited to the wrong player and corrected later. They get cancelled after review. Sometimes a goal appears in the feed even though it never happened because play had already stopped.

Penalty shootouts are probably the hardest case: many events in a few minutes, with each one able to change the result.

**Giving our development agents a short brief.** We expected the agents to know enough football on their own. They did not.

We had to spell out the rules in detail, especially around added time. A match can stop for an injury and run for another 20 minutes. It can be postponed because of weather and finish the next day.

No guardrail catches every inconsistency in live data. The agents need those rules explicitly written down if they are going to explain what they see correctly.

## The numbers so far

**The app in numbers, February to October 2026**

Release count and rating from the app’s App Store Connect dashboard.

|                              |                                        |
| ---------------------------- | -------------------------------------- |
| Concept to launch            | Under 6 months (February to June 2026) |
| App Store releases           | 22 (8 June to 1 October 2026)          |
| Release cadence              | Weekly or faster                       |
| App Store rating             | 4.8 (36 ratings, October 2026)         |
| ML models integrated         | About 40                               |
| Competitions covered         | MLS, Premier League, World Cup         |
| Phases of play measured      | More than 10                           |
| Shareable visual story types | 5                                      |

*The app isn’t named at the client’s request.*

## My take

Looking back, integrating the models was probably the more predictable part of the project.

The harder work was everything that had to be true before their output could be trusted: the correct events, the correct score, the timing and all the strange edge cases that appear in live football.

We also shipped much more frequently than we originally expected. Between June and October we released 22 versions of the app. A lot of what we know now came from watching real matches, seeing something strange in production and fixing it before the next round of games.

AI helped us move much faster, both in development and in turning match data into easily readable content. My estimate is that a comparable app would have taken us around 1–2 years before agents. But it did not remove the need for football knowledge or software development knowledge or QA testing.

For example, if the implementation of extra-time rules misses an edge case and causes a critical bug, the development agent might give a very convincing but incorrect explanation, pointing to a data issue instead of a code bug. This makes industry knowledge even more important than it used to be.

### What I think is transferable

A few lessons here apply beyond football.

The first is building against a deadline that cannot move. The second is working with live data that may be incomplete or change after you receive it. The third is knowing the industry well enough to spot when development agents are heading in the wrong direction, rather than expecting them to identify and fix every underlying problem on their own.

We see the same pattern in other products where timing and live information matter.

For our client, the World Cup made all three happen at once.

More on football from the lab: [a cash forecast for football clubs](https://x2labs.ai/notes/football-club-cash-forecast), built from contract terms and payment history, and [a travel dashboard](https://x2labs.ai/notes/football-team-travel-planner) that shows what breaks when a fixture moves.

**Notes cited**

- [Cash forecasts for football clubs: 18% of contract terms fixed by hand](https://x2labs.ai/notes/football-club-cash-forecast) · Dan Mocan · 19 Oct 2026
- [When a fixture moves, what breaks? A football travel dashboard](https://x2labs.ai/notes/football-team-travel-planner) · Dan Mocan · 2 Nov 2026
