Novessa Pitchside
Microsoft x Premier League "Inside the Game" developer hackathon
Microsoft Agent Framework, Azure AI Foundry, azure-ai-evaluation, Cosmos DB, Container Apps, Event Hubs, Bicep, azd, GitHub Actions, MCP, A2A

Why
Everyone watching a match gets the same broadcast. Novessa Pitchside gives each viewer their own version of the game, live: an analyst gets the pressing numbers, a casual fan gets the story, in their language, built around their club or player.
What
A personalised live football broadcast, built for the Microsoft x Premier League hackathon. A team of AI agents decides which moments matter, explains them for each viewer in their language, and proves every number before it airs.
How
A Producer, an Analyst, a Verifier, a Narrator and a Gate, built in the Microsoft Agent Framework. They run on two Foundry deployments, a small chat model and a reasoning model. Output is scored with azure-ai-evaluation and the verified result is cached in Cosmos DB. The API, my own MCP server and an A2A endpoint ship as three Container Apps from one image, scaling from 1 to 3 replicas.
The hard part
Trust. A model never computes a number: a stats layer does, in code, and nothing a model writes reaches the screen until the Verifier has recomputed every figure. When a claim fails, the viewer gets the plain numbers instead of a nice-sounding guess, and any agent can be killed mid-match while the broadcast carries on.
Result
On the same 41 moments, a single prompt aired 10 story errors and 3 unsupported numbers; the agent team aired 0 and 0. 150 simulated viewers made 32,120 requests with 0 errors at a p95 of 11 ms, and Event Hubs carried 19,830 messages with none lost. 262 Python tests, 88 frontend unit tests and 16 end-to-end tests. Deployed with Bicep, azd and GitHub Actions, using Managed Identity so there are no keys.
Next projectGulfSafe