Sports intelligence · Production

WhoWins.AI

Compare pregame model probabilities, market context, and community picks without hiding uncertainty.

WhoWins.AI combines a public web product, native iOS experience, production API, sportsbook data pipeline, and sport-specific model research in one product workspace.

Explore the case study Website

App Store

Category: Sports intelligence
Status: Production
Platforms: Web · iPhone
Built with: Next.js · SwiftUI · Firebase

The product problem

Why WhoWins.AI exists.

Prediction products often collapse a probability, an explanation, and a marketing claim into one confident number. WhoWins separates those jobs so disagreement and missing evidence remain visible.

What the product does

The useful parts, stated plainly.

These capabilities are grounded in the project’s README, metadata, source structure, and checked-in product documentation.

Browse upcoming matchups

Move through games, teams, and matchup detail with current model and market context.

Compare probabilities

Keep the Hermes probability layer distinct from the written matchup explanation.

Inspect the factors

A structured rubric exposes market, team strength, injuries, venue, schedule, fit, and uncertainty.

Read the board

Normalize sportsbook prices into comparable no-vig probabilities and surface line context where data is available.

Make a community pick

Profiles, leaderboards, comments, and prediction history create a social layer around the research.

Follow the methodology

Dedicated research pages document questions, out-of-sample evaluation, baselines, limitations, and product implications.

Designed for

People with a specific job to do.

Sports fans comparing a matchup
Model-curious readers
People learning how odds encode probability
Communities making and tracking picks

Product choices

What makes it distinctly Orion.

The workspace separates production web/API code from model research and pipeline development.
Model probabilities and natural-language scorecards are distinct layers.
Research compares models against the betting market rather than only an easy home-team baseline.

Product gallery

No invented screens.

The repositories contain polished App Store campaign compositions, not raw app screens. They were intentionally excluded from the device gallery; exact iPhone and web capture requests are documented in SCREENSHOT_REQUESTS.md.

Capture request prepared

Nothing misleading has been substituted. The requested device, state, demo data, native resolution, filenames, and repository destination are documented for the owner.

WhoWins.AI research

The method and the limitation belong together.

Public-safe summaries from the project’s substantive research and product documentation.

July 2026

Football probabilities evaluated against the market

Question: Can leakage-safe team and market features produce useful NFL and college-football probabilities when evaluated season by season?
Method: The documented Hermes study uses pregame-only feature tables, trains through season N−1, tests on season N, repeats across multiple seasons, and includes home-team, Elo, and no-vig closing-market baselines.
Key finding: The important result is not a guarantee of beating closing markets. The documented NFL stack lands near the closing-line accuracy baseline while the stats-only view showed a placebo-controlled relationship with later line movement.
Limitation: The models are pregame winner models only. Market coverage varies, early-season features lean on prior team state, and no reported result removes uncertainty or forms a betting strategy.
Product implication: The product exposes stats-only, market, and blended views so agreement and disagreement can be inspected instead of reduced to a “lock.”

May 2026

When a smaller market model beats a larger feature set

Question: How much do team history, recent form, sequence models, and sportsbook context each contribute to NBA pregame probability?
Method: The NBA work built leakage-safe rolling team features, compared tabular and sequence models, then enriched the odds era with no-vig probability, lines, bookmaker counts, timing, and movement.
Key finding: In the documented split, compact market-aware candidates were more stable than the larger basketball-only feature set. That is a finding about the current sample, not evidence that team context is irrelevant.
Limitation: Historical odds begin later than the game dataset, the research snapshot omits the newest seasons, and rolling-origin retesting remains important.
Product implication: Hermes treats market consensus as a serious input while preserving a separate basketball layer for future residual modeling.