I’ve just launched V1 of a Fantasy Premier League (FPL) insights app!

What does it do? Pulls insights from the many, many content creators out there, and classifies them. Then once they’ve been classified it saves them in a database, which you can query using AI to generate weekly reports, team-specific reports, and ask any kind of questions about the collective wisdom of the FPL hive mind that you might wish.

The idea is to get all of that great advice without sitting through interminable podasts. And to make better FPL decisions, never missing potential opportunities or big movements in the game that, the content creators are flagging.

Want to cut right to the chase? you can see a read only version of the database here: https://fpl.john-platt.com/

Or for the full version, download the Github repo and follow the instrutions.

If you want to know more abot how, and why I did this, and a little about how it works, read on below.

What’s an FPL, anyway?

For the uninitiated, FPL is Fantasy Premier League. And if you’re among the 12 million currently active players (my commiserations), you can skip to the solution section below, and save your precious time to consider whether Calvert-Lewin or Thierno Barry are better value in the £6 million striker bracket (it’s Barry).

Good, for those of you still here, FPL is a game in which you choose 15 real Premier League players, with a fake budget of £100 million. Based on their real performance (goals, assists and clean sheets, mostly), they get you points throughout the season.

If you really, really want to know the rules, you can click here. But otherwise, that’s about enough to get the gist of this.

The important thing is that I have a league with my friends, running for the last 9 years, which I’ve only won once. Technically it was also a joint win, so this year I’m absolutely determined to win it outright.

The prize; a slap-up meal with the group at Madrid’s Casa Juan, celebratory restaurant of choice for the Real Madrid football team, and monetarily nowhere near worth the tens of hours put into playing the game each year.

What I’m trying to say is, we play for pride.

Screenshot of a gaming dashboard displaying user 'John Platt' under the username 'Palmersan Cheese'. It shows overall points as 607, overall rank as 403,444, total players as 11,232,225, and gameweek points as 74.
My current ranking, putting me first in the Big Feed IX

The problem

Each week you can make one transfer. Making more than one costs you 4 points, so knowing who is playing well, who isn’t, and who is going to go on a favourable run of games, is essential.

Because of this, a thriving niche content industry has sprung up, of (usually) bearded men on live streams, poring over stats to give recommendations to their followers. Fantasy Football Scout, Fantasy Football Hub, and more, charge a subscription fee for their recommendations. All focussed on helping you beat out your friends to a mini league victory.

It’s a content cottage industry worthy of study (and one that I may dig into in the future).

But the main problem is that these people are truly uncharismatic. They know a lot about the Premier League, the statistics behind it, and generally make good decisions. But the overlap between these kind of profiles and charismatic podcasters is limited.

The videos are also primarily live stream replays, meaning you have to sit through all of the awkward umming, ahhing and small talk as the hosts wait for viewers to join. It’s not unheard of for one of the hosts to leave for 10 minutes because someone came to the door, and they have a rambling, unedited feel.

The solution I came up with

I was spending quite a bit of time listening to these podcasts for their insights. But sitting through hours of them for a handful of useful recommendations was getting frustrating.

Is there not some way that we can use AI, avoid these mind-numbing listens, and leverage the wisdom of the crowds to come up with optimal FPL decisions – winning a slap up meal in then process?

I wanted some kind of app that could take all of the content being produced by these FPL experts, extract the useful information, and put it all in one place. To benefit from their knowledge, without having to… actually listen to them.

Who was playing well, who wasn’t. Which players were undervalued, who was likely to be dropped, which teams had good fixtures, and the best ‘chip strategies’ (That needs no explanation if you’re an FPL nerd, and if you’re not, you can call them ‘jokers’)

Four gameplay options for fantasy football: Bench Boost, Triple Captain, Wildcard, and Free Hit, each with a 'Play' button.

But rather than a boring database of transcripts, what interested me was the idea of building up a sort of collective intelligence from all of these people. Opinions, predictions and recommendations that would grow throughout the season.

Why not just ask Claude for advice?

And it’s something I can only get from humans. Because a lot of what makes someone good at FPL is difficult to capture in statistics alone.

You can look up how many goals a player has scored, or how many shots he’s taken, easily enough. But knowing that he’s looked particularly threatening in the last three games, that his manager has changed the formation to suit him, or that he was unlucky not to score three last weekend, requires someone to have actually watched the games and understand football.

A footballer in a blue Everton kit looks frustrated, with his hands on his head, during a match in rainy weather.
Thierno Barry, definition of a man ‘unlucky’ not to score three

My thinking was that if I could somehow get all of this information into a database, I could give an AI access to the sort of context it would otherwise be missing.

And then, ideally, combine all that knowledge with information about my own team. The players I already owned, how much money I had in the bank, what transfers I was considering, and so on.

Rather than asking ChatGPT something generic like ‘who are the best strikers in FPL?’, I could ask it what I should actually do with my team this week, and get an answer based on the opinion of a collection way too immersed in fantasy football.

The wisdom of the crowds, essentially.

How the AI works

Having decided what I wanted, the next question was how to actually build it. For this, I used Codex, OpenAI’s coding tool. I explained what I wanted the app to do, and we started building it.

Underneath, it’s a Next.js app, written in TypeScript, with a Supabase database that stores all of the information, and the OpenAI API doing most of the work involving language.

There are essentially four parts:

Transcript extraction

Using Supadata, with a local Python extractor as a fallback, we extract transcripts from selected content creators, or by manually inserting relevant URLs.

Screenshot of the Personal FPL Knowledge Base dashboard showing sources and processing information, including source statistics and processing status.

Insight categorisation

The next step is using the OpenAI API to categorise the insights within each transcript, and associate them with teams, individual players and the creator themself. These are matched against the official FPL player database, so the insights are associated with all the relevant data (price, points, upcoming fixtures, team, whether you own them or not…).

So if a creator says that they think Arsenal have an excellent run of fixtures coming up, it would be associated with Arsenal. If the creator thinks Saka is going to go on a goalscoring run, it would be associated with both Arsenal and Saka.

Both of those insights would also be associated with the creator who said them. For the moment, creator association isn’t especially useful, apart from being able to browse through what each creator has said, if needed. But in the future, I can imagine introducing a weighting system to prioritise creators who have historically been correct more often in their predictions.

Keeping these insights in their atomised form means each one can be used in different contexts. Rather than processing a summary of the whole transcript every time, we can reuse individual insights, saving on API tokens.

FPL Gameweek 10 preview featuring a discussion on defensive players, focusing on Arsenal's Gabriel and other key defenders for the Fantasy Premier League 2026/27.

Generating weekly briefings

Screenshot of a web interface for generating a research briefing, with fields for date range and topic selection, and a loading message indicating the briefing is being generated.

Games come week in, week out. We’ll need a weekly briefing. The weekly briefing uses a different OpenAI API run and model – and all of the creator insights from a given period – to synthesise the most commonly recurring topics.

A research briefing document discussing Premier League football analysis, focusing on teams like Arsenal, Bournemouth, and Brighton, along with player insights for Fantasy Premier League strategies.

Those topics are clustered by player and team, and across key strategic decisions.

There’s one important detail here, and that’s ‘split decisions’. A split decision is when opinion varies across multiple creators, something that happens on a regular basis. One creator recommends selling Saka because he’s overpriced. Another recommends holding, due to his fixtures.

These are highlighted separately in the briefing as key decisions to be made that week, then stored in the database for retrieval further down the line.

A claim records what one creator thinks. A consensus is when multiple creators say the same thing, and we should pay serious attention to that. A split decision records what the manager has to weigh when creators disagree.

All this comes out in a weekly briefing.

Screenshot of a fantasy football discussion forum highlighting key players, Gabriel Magalhães and Erling Haaland, with commentary on their performance and value.

Ask the crowd

Screenshot of a web interface asking for research about Fantasy Premier League (FPL), featuring fields for user questions, research time scope, and context.

This, then, is where the magic happens. We have a database filled with insights on players, on teams and on key decisions that need to be made this week.

We have also uploaded our own teams, so stored in the database is how much money we have to spend, who is in our team, their fixtures and form.

So we can query that database, and use the AI model to reason and respond to questions, either with the context of our team, or more generally.

I can ask:

‘Should I be selling Saka this week?’

The app uses AI to interpret the question, works out which information is relevant, and then searches the database for information on my team, Saka, and the claims and disagreements that might help answer it.

It then uses AI again to put together a response, drawing on the evidence it’s found. So instead of getting a generic answer based on whatever ChatGPT happens to know about football, I’m getting an answer based on what actual creators have been saying, and the reasoning behind their arguments.

Screenshot of a webpage titled 'Ask anything about FPL', featuring a question on Fantasy Premier League transfers and advice from creators.

Technically, this is retrieval-augmented generation, or RAG: a way of giving an AI access to a specific body of knowledge, rather than expecting it to know everything itself.

In my case, the knowledge just happens to be a collection of opinions about fantasy football.

Trying out the web app

Want to give it a try? There are two ways you can do so.

1: I’ve deployed a read-only V1 to fpl.john-platt.com. It has a similar interface, with all the insights from different creators and my own generated briefings, all read-only.

2: For the full experience, the GitHub is public, and you can download it to run on your own machine, using your own API key and Supabase database.

Why haven’t I deployed the full experience? Extracting insights and running analyses requires an OpenAI API key. And while it’s cheap enough to run for myself, it could become a money sink if anyone could run analyses using my key.

Right now I’m not looking to monetise the app, so I’d rather give it away for you to run locally with your own key, if interested.

Next steps

There are still a few things I’d like to add.

For a start, I’d like to give the app more of the statistical context behind each player. Heat maps, form, and other performance data, which could sit alongside the creator insights on individual player and team pages, and be used by the AI when answering questions.

Then there’s the FPL-specific information. Things like price changes, which can make a fairly substantial difference to when you should make a transfer. Particularly annoying when you find yourself £0.1 imaginary million short of the perfect team.

But perhaps the most interesting next step would be to bring in information about my mini league. Where I stand against my friends, which players they own, and who I need to catch 😈

Making the best decision in FPL isn’t always the same as making the best decision to win your mini league. If you’re 50 points behind in the end game, you need to take more risks than if you’re comfortably ahead.

And that’s ultimately where I’d like to get to. An app that doesn’t just know what a load of FPL experts are thinking, but can use that knowledge, the statistics, and my own circumstances to help me make better decisions.

All in pursuit of a free dinner which, at this point, definitely isn’t going to cover the time I’ve spent trying to win it.