A live voice agent has no margin for latency: if the commentary arrives after the goal, the product is worthless, so everything was built backwards from that constraint.
2
App stores, live
iOS and Android from one codebase
8
Leagues and tournaments
through to the 2026 World Cup
7
Languages at launch
commentary and interface
The problem
Football is a social product consumed, increasingly, alone. The gap is not information; scores are everywhere. What is missing is reaction, and somebody to shout at when the penalty is given.
An AI commentator only fills that gap if it is genuinely live. Commentary that lags the match by even a few seconds stops being a companion and becomes a transcript, and a transcript is worth nothing when the viewer already saw the goal.
What we did
Victoria is a real-time voice agent rather than a text feed with speech bolted on. LiveKit streams the audio at low latency, ElevenLabs handles speech in and out, and an LLM writes commentary against live match state across the Premier League, La Liga, Bundesliga, Serie A, Ligue 1, the Champions League, MLS and the 2026 World Cup.
It is two-way. You can cut in by voice or by chat mid-match and she answers, which is the difference between a broadcast and a companion.
One React Native and Expo codebase ships to iOS and Android. A Node backend on AWS runs predictions, the credit system and affiliate rewards.
There is no subscription. Credits are bought in packs when you want them, which suits a product whose usage spikes around fixtures rather than running evenly across a month.
The outcome
The app is published on the App Store and Google Play, covering eight competitions in seven languages, with the voice agent, credit system and affiliate rewards all running in production.
Mobile
Real-time voice
Backend
Commerce
What we learned
For a live product, latency is the feature. Every architectural decision here follows from the commentary needing to land before the viewer's own reaction does.
Match the billing model to the usage pattern. Spiky, event-driven usage and monthly subscriptions are a bad fit, and the churn tells you so.
One codebase to two stores is still the right default for a product that has to prove itself before it can afford two native teams.
Thirty minutes with the engineer who would lead the project. You will leave with an opinion on what to build, what to skip, and roughly what it costs, whether or not you hire us.
2 build slots open · Next start: early next month
What happens next