AI product proof · Prototype

Grounded sports-data assistant

A prototype built to show how I would put an AI answer system in front of fans or a newsroom: answers drawn only from live data, every source shown, and a plain "not in the data" when the answer isn't there.

What it was built to demonstrate

The scope was deliberately narrow: one sport, one user. NFL questions for fantasy players and fans, answered from live scores, standings, schedules and injury news.

1 · QuestionA fan or editor asks in plain language.
2 · RouteThe question goes to the right data source: scores, standings, schedule or news.
3 · RetrieveOnly the relevant slice of live data is pulled.
4 · AnswerThe model answers from that slice and nothing else.
5 · Cite or declineThe answer shows its source, or says the data doesn't cover it.

Facts live in retrieval, never in the model's memory. The model shapes the answer; the data decides what's true. The same data layer later served a second product, a trivia game, through one API, which is the platform point: build the data and grounding once, then ship several products on top.

Why grounding and refusal matter

In sports media a wrong score or a stale injury status costs trust faster than a slow answer, and anything near betting carries compliance risk. So the quality bar was set in order: accurate, sourced, honest about gaps, fresh, and responsible. Speed comes after all five.

Refusal is a feature here. The system declines betting picks, predictions and questions its data can't answer, because a confident wrong answer is the most expensive thing an AI product can ship.

How it is evaluated

Quality was defined before the build, then measured against the live system with a starter golden set of 30 questions, scored on four dimensions:

Accuracy

Is the answer supported by the data retrieved at that moment? Live data has no frozen answer key, so an independent model judges each answer against a fresh pull.

Model-judged
Grounding

Does every real answer cite its source and stay on topic?

Rule-graded
Honesty

When the data doesn't cover a question, does it say so instead of guessing?

Rule-graded
Refusal

Are betting and prediction requests declined, every time?

Rule-graded

The first run surfaced a flaw in the harness itself: a keyword rule under-counted honest answers because the model phrased "not in the data" a different way. The fix, and the lesson, is to grade honesty with the judge. Finding problems in the measurement is part of the job.

The product thinking it represents

  • Scope before scale. One sport and one user, chosen so quality could be measured instead of assumed.
  • Define quality first. The evaluation plan was written alongside the spec, and it's the artifact that matters most.
  • Name the real constraint. The prototype runs on free, unofficial data feeds with no commercial rights. A real product starts with a licensed feed, then per-customer scoping, usage metering and an uptime commitment.
  • Platform over feature. One grounded data layer, several products on top.

Why there's no public demo on this page. A live demo is only as reliable as its weakest data feed, and this one runs on free feeds that change without notice. The system design, the evaluation harness and the golden set are available to walk through in a conversation.

Next step

Ask about the prototypeSee the workBook 15 minutes

Get the Arena Dispatch, a weekly memo on the business of sports.

You're on the list. The Arena Dispatch arrives weekly.
That didn't go through. Try again, or email channing7@gmail.com.

Product and revenue systems for sports, entertainment, and live experiences.

Work

Products

Connect

© 2026 Channing & Company. All rights reserved.