The AI sports commentary engine.
Built from the snap up.
gamedai is the first AI-generated live sports radio broadcast. The radio reads from a knowledge graph built from live snaps and historical data, not a generic language model. Every call lands 15 to 25 seconds after the event in your ear. Three interactive surfaces: snap-level narration, voice questions, and fantasy reactions. Hands free. Eyes free. In the game.
Three steps. No setup.
Open the app.
Pick the voice. Studio Analyst. Coach Bill. Fantasy Analyst. Casual. Four products in one. Same game, four ways to hear it.
Hit play. Hands free.
It’s a broadcast, not a chatbot. The radio calls every snap. You listen on whatever you’ve already got on. Headphones. Earbuds. Smart glasses. The phone stays where it is.
Ask anything.
“What’s their record on third-and-short?” “Should they have gone for it on fourth?” “Why isn’t my tight end getting targets?” Hold the button. Ask. Get the answer in the voice you picked.
5-layer architecture.
Engineer-built. Coach-edited.
Layer 1: NeSy ground truth — deterministic event detection on every snap. Layer 2: validated ontology — facts ratified against schema before they reach narration. Layer 3: RAG retrieval — the knowledge graph pulls in matching prior plays. Layer 4: Claude Sonnet 4.6 — voice rendering in the persona you picked. Layer 5: an eval filter — every utterance is gate-checked before it goes to your ear.
Data sources: nflfastR play-by-play, PFF Premium charting, Big Data Bowl 2026 historical patterns. Source-cite-able beats vague.
Read more about the AI football analyst voices calling the games, or when AI radio NFL preseason coverage goes live.
The platform
underneath.
One engine, many surfaces.
The radio is what you hear. Underneath sits a five-layer engine: deterministic event detection, multi-source ontology, per-turn retrieval, persona narration, eval filter. Every surface — radio, the AI butler, the talk shows, the analytics tools, the future creator stations — runs on the same ground truth.
The ontology is the moat. Every source attributed. Every claim multi-source merged. Every interpretation jury-validated. Same engine, every sport.
Every signal, measured end-to-end.
Every play, every question, every fantasy event flows through the same five layers — and we watch all of it. Per-layer latency. Per-component health. Per-surface usage. Bottleneck attribution. Cost-per-utterance.
When a signal slows down, we know which layer owns it. When a component degrades, we know inside fifteen minutes. When the answer arrives in your headphones, the trace from event to audio is one ID we can follow.
Built for operators, not for screenshots.