Football stories don't start and end with the scoreline. We’re excited to celebrate the six winning teams, from five countries of Inside the Game: Developer Hackathon, presented by Microsoft and the Premier League. The winners were announced at Microsoft Ignite 2026.
During Inside the Game, developers built AI apps and agents that turn synthetic match data into real-time insights, explainable match intelligence and personalized recaps for every fan. They built them with Azure, GitHub, GitHub Copilot, Microsoft Foundry and Visual Studio Code.
The strongest entries told real football stories, backed up every number on screen and were ready for real studio and streaming workflows. Congratulations to every team and thank you for spending your evenings and weekends inside the game with us.
🏆 The winning projects
Grand Prize, First Place: Pitch Ledger – GitHub Repo
Built by Ephraim Cabanlit – Philippines
Pitch Ledger starts with a clear principle: let AI tell the story, but let code keep the score. It tackles one of the biggest barriers to putting generative AI on air—a made-up statistic. One wrong number can cost more trust than a hundred right ones earn. Here’s how the project puts that principle into practice. Synthetic match events stream through Azure Event Hubs into a deterministic engine: code, rather than a language model, calculates possession, pressing, tempo and player physical metrics. The engine flags key moments, including momentum shifts, goals and red cards, and attaches the evidence behind each one. Five agents built with Microsoft Agent Framework on Microsoft Foundry turn those moments into stories. The Explainer drafts why a moment matters, using a read-only MCP stats server to access the numbers without creating or changing them. The Auditor independently recalculates every figure from the raw events, sending drafts back for correction or blocking them when they go beyond the evidence.
Pitch Ledger makes that approach testable. Its Match Lab simulates full matches between six fictional clubs, generating about 125,000 events and tracking rows per match. It plants known scenarios—a late collapse or a switch to a high press—so each demo match becomes a test the engine must pass. In the demo match, the engine flagged 54 moments; 17 aired after audit and producer approval, with about US$0.07 in model cost for the whole match. When a false statistic was deliberately planted, the Auditor caught it before it could air. Every insight is recorded in a ledger with its numbers, supporting evidence and full audit trail, making it possible to trace what reached the screen and why. The solution runs keyless on Azure, using managed identities instead of stored credentials, and scales to zero when idle.
Grand Prize, Second Place: Footballytics – GitHub Repo
Built by Farquleet Farhat Gondal, Fahad Ramzan, Sheheryar Ramzan – Pakistan
Your club is playing, your fantasy captain is on another pitch, and something is happening in a third match. Where do you look? Footballytics puts a Personal Director agent in charge of that decision, following your interests across ten simultaneous matches and choosing when to switch the action or open picture-in-picture. Every card explains why it’s on your screen, and the experience learns from what you skip, like and replay. The result is a broadcast shaped not just by what happens on the pitch, but by what matters to the person watching.
To get started, choose a club, how you like to watch and one statistic to follow. The app builds a fantasy squad and tracks it across all ten matches. Your squad—with your captain counting double—alongside your club, favorite statistic and previous likes guides the Personal Director’s choices. Tap a player and Player Cam follows them, while a Taste Learner uses your interactions to keep adapting the experience. Stories come in fan, analyst, kids and audio-description modes, written natively in seven languages.
Behind that experience is an agent newsroom built on Microsoft Agent Framework and Microsoft Foundry, with a clear division of responsibility: code calculates the facts, and agents tell the story. A deterministic stats engine calculates pass difficulty, pitch control, momentum, pressing and live fantasy points. An Analyst agent explains why each moment matters, while a deterministic Verifier checks every claim against live data and sends incorrect numbers back for a rewrite. A Host agent then writes the verified story for each audience and language. The team makes that safeguard testable: viewers can inflate a statistic by 50% or invent a number in the Studio and watch the Verifier reject it.
The team also designed for the practical demands of a live broadcast. Template captions appear immediately, with agent-written lines replacing them about 10 seconds later. If the models go down, the broadcast falls back to templates instead of going blank. When nobody is watching, the matchday pauses, so agents only spend tokens when there’s an audience. A Football Intelligence MCP server exposes the same 12 tools the agents use, letting developers query the live matchday from GitHub Copilot agent mode.
Best Use of Microsoft Foundry: Matchcut – GitHub Repo
Built by Malik Bello – Nigeria
Matchcut promises every fan their own cut of the game. It turns a live stream of synthetic football events into verified, explainable insight presented in a broadcast-style 3D show. Minutes after the final whistle, it gives every fan a voiced, captioned highlights film of their club, their player and their language. Films come in seven languages. They're formatted for phones and TVs and can be shared by link or QR code. Every film also has an audio-described version for blind and partially sighted fans. A describer explains what's on screen in the gaps the commentary leaves: who has the ball, the build-up, the goal, the celebration and each replay.
During the match, agents on Microsoft Foundry run as a Microsoft Agent Framework workflow. An Analyst and a Historian work in parallel, a Writer drafts the line, and a deterministic Verifier checks every number against the fact it cites. Nothing unverified airs. A League Memory server gives the agents five seasons of league history through MCP, so a goal arrives with its story, such as a striker's ninth of the season and his second of the night. At full time, the Recap Factory sends one queue message per fan to Azure Container Apps Jobs, which scale from zero to 20 replicas. In one measured run, it produced 42 films in just over three minutes at about two cents per film. Before any fan sees a film, a QA agent watches and listens to it: Azure AI Speech transcribes the audio, a model reads the captions off the frames, and a loudness meter checks the sound.
Best Enterprise Solution: Absolute Scenes – GitHub Repo
Built by Manu Areraa A K – India
Absolute Scenes is an AI pundit that draws as it talks. It turns match tracking data into a studio-style breakdown. A crew of agents picks the key moment, measures it with analytics tools, then writes and voices a short script. Each arrow, line and zone appears on the pitch at the exact word that names it, and every number on screen links back to the tool call that measured it. Viewers can also ask the pundit about any moment, typed or spoken. An Azure-based agent crew responds to typed or spoken questions with verified answers, drawing its explanation directly on the pitch. For "what should they have done?" questions, a Tactician agent proposes new player positions, and the simulator replays the moment. A Critic agent keeps the proposed change only if it improves the simulated outcome. Fans can even drag the players themselves and see whether they beat the Tactician.
Behind the show is a clear rule: facts come from tools, never from the model. Deterministic analytics tools calculate distances, gaps and pitch control. The language model only picks, phrases and orders the story, and a Verifier agent re-runs the tools before anything goes on air. A producer console lets an editor approve or reject each moment before it airs. Prompt Shields screen every viewer question, and the show adapts to 18 versions across audience, supporter side and language (English and Spanish).
Best Multi-Agent System: MatchCast – GitHub Repo
Built by Akshay Pai – India
MatchCast is an AI production crew for a live football broadcast, with a separate agent for each job a broadcast team would do. A producer picks the moment, an Analyst finds the numbers and turns match events into claims that cite their evidence, a Narrator writes the commentary line and a Verifier checks it before air by recounting every number from the raw event log.
Crucially, the Verifier is never the agent that wrote the line. If a check fails, the work goes back to the agent responsible: a wrong number to the Analyst, wording to the Narrator. The agents run as a typed Microsoft Agent Framework graph on Microsoft Foundry. When a goal or red card calls for deeper coverage, a planner sends up to five specialist agents to work in parallel on replay, tactics, stats, rhythm and head-to-head history. A partner Stats Desk agent joins over A2A.
Nothing reaches the screen until the Verifier has recounted it. Lines that pass are screened by Azure AI Content Safety, and booth commentary waits for a producer to approve, edit or kill it in the Production Studio. A tamper-evident audit chain records every step, and a “Why?” button on any line shows the claim, the events it cites and each agent’s work. Fans can switch between fan and analyst commentary, pick from six feeds including audio description and a half-time explainer, and read lines in ten languages through Azure AI Translator.
Best Azure Cloud Native Integration: Talkback – GitHub Repo
Built by Yehor Ivashchenko – Poland
The stat was right. The story was missing. A correct number can still leave fans without the context that makes a moment matter. Talkback focuses on that editorial challenge: finding the story, matching the strength of the wording to the evidence, and adapting it for different markets. Its AI editorial gallery is a workspace for an entire matchday, where specialized agents uncover overlooked stories, check the evidence behind each line and clear it separately for every territory where it will air.
The agents tackle different parts of that editorial work. A Spotter ranks moments across ten pitches, an Archivist, running as a Foundry hosted agent, add context by finding storylines such as a player sitting on 99 career goals. A Lab agent replays a moment 120 times to test whether the evidence supports casual wording such as "because," rather than simply describing what happened.
The tech behind the winners
Across the winning projects, Microsoft Foundry and Microsoft Agent Framework were the core of the stack, running multi-agent workflows and Foundry hosted agents. Teams streamed live match events through Azure Event Hubs and ran their agents and services on Azure Container Apps. They stored and queried data in Azure Database Services, Azure Cosmos DB, Azure Data Explorer and Azure Blob Storage. MCP servers gave agents and GitHub Copilot agent mode the same football stats tools, and A2A support let agents work together. Azure AI Speech produced voiced, audio-described highlight films, and Azure AI Content Safety checked on-air wording. Application Insights with Foundry Tracing and the Azure AI Evaluation SDK let teams monitor and evaluate their agents. Most solutions ran keyless, secured with managed identities and Microsoft Entra ID.
What these projects have in common
Look across the six winners and a clear pattern emerges. Almost every team separated the creative work of language models from the deterministic work of checking facts, so a verifier or auditor stands between an AI-written line and the audience. They treated multilingual, accessible personalization as a core feature rather than an add-on. And they built on cloud-native Azure architectures that could run in a real production environment. These winners show what happens when AI is built for both impact and trust: football intelligence that is explainable, personalized and ready for real-world production at global scale.
Thank you
Thank you to every developer who registered, built and submitted a project. Your work set a high bar. We are also grateful to the Premier League for partnering with us and lending their expertise to the judging, and to the Microsoft judges, Cloud Advocates, MVPs, Student Ambassadors and community leaders who mentored, promoted and supported the hackathon along the way.
Keep building
The winning repositories are public, so explore them, learn from their architectures. If you want to start your own match intelligence project, the Inside the Game livestream series on Microsoft Reactor is available on demand: Inside the Game: Developer Hackathon on Microsoft Reactor.
Congratulations again to all our winners. We can't wait to see what you build next.
Microsoft Global Community Hackathon Team