AR Basketball Statistics

The problem

Coaches watch live games and need instant answers — how many turnovers, effective field goal % against this defense, plus-minus for the current lineup — without pulling out a laptop mid-timeout. Existing analytics tools bury the data three menus deep, so most of it never gets used at the moment it would actually influence a decision. The existing product landscape splits into two camps: dashboards (too slow, too dense for in-game use) and simple stat overlays (too shallow for real decisions). Neither serves the coach who wants to ask a question the way they'd ask an assistant.

Key challenges

Live game data is messy — stat feeds have latency, occasional errors, and format inconsistencies across venues and leagues. Voice interfaces in a loud arena face acoustic challenges most consumer voice AI doesn't handle. Query interpretation has to work despite domain jargon and personal shorthand ("what's their EFG on switches" is a valid coach question). And the system has to work offline for venues with unreliable connectivity, which changes every architectural assumption downstream.

What we built

AR Data built a voice-first analytics layer over live game data — coaches ask questions aloud, the system pulls from real-time stat feeds and prior-game history, then responds with the number and the context. The interpretation layer is trained on basketball-specific vocabulary and coach shorthand. The retrieval layer aggregates across current game, season history, and opponent scouting data. Runs offline for arenas with poor connectivity: a local cache of the current game state and recent opponent data lives on-device, with sync-back when connectivity returns. The response format is optimized for the coach's context — short and specific during a timeout, more contextual during timeouts or post-game.

Our approach

1. 1

   Voice interface designed for arena acoustics

   The audio pipeline uses noise-adapted STT models and directional-mic capture profiles tuned for courtside environments. This is what makes the system work in the actual conditions of a live game rather than only in a demo.

2. 2

   Basketball vocabulary and shorthand in the interpretation layer

   Query interpretation is grounded in basketball-specific vocabulary (EFG, PPP, defensive rating, ATO) and coach shorthand. Generic language models struggle with the terminology; domain grounding fixes it.

3. 3

   Offline-first data model

   Current game state and opponent scouting data live on-device with sync-back. Arenas with poor connectivity are the norm, not the exception, and the system has to work through them.

4. 4

   Context-aware response formatting

   During live play: short and specific. In timeouts: more context. Post-game: full narrative. The same query produces different responses depending on the coach's current context.

Key architectural decisions

Results

• Voice-driven live stat queries during play, timeouts, and post-game
• Real-time analytics with sub-second response inside timeouts
• Historical opponent context available on demand
• Offline-capable — works through arena connectivity gaps
• Domain-tuned voice interpretation for basketball vocabulary
• Coach-friendly response formatting — short in-play, contextual otherwise
• Session sync-back preserves full record for post-game review

Impact

AR Basketball Statistics is a working example of voice AI applied to a domain the mainstream voice ecosystem ignores. The engagement produced the pattern we now bring to any voice-in-domain project: match the acoustic environment first, ground interpretation in domain vocabulary, and design for offline as the primary case. The coaches who use it don't describe it as "voice AI" — they describe it as "asking my analyst," which is the framing that means the interface disappeared.

Tech stack