Football Market Decision Lab
An explainable, memory-aware review queue for fast-moving football markets, demonstrated with synthetic data and no wagering execution.
What this project proves
AI agents support isolated implementation, testing, and pull requests; I own the data model, scoring rationale, human-approval boundaries, and public sanitisation.
A deployed synthetic-data demo showing 90-minute signal memory, reasoned candidates, and human review, with no real account, wagering interface, or performance claim.
90 minActive signal-memory window
2-stepHuman merge approval
0Automated wagering actions
Problem
Market updates arrive quickly and lose context. Latest values alone cannot explain whether a move persists, why it matters, or whether it deserves human review.
Build
Built low-frequency collection, normalisation, 90-minute signal memory, quantitative scoring, and human review, supported by a constrained AI-agent workflow for tests and pull requests.
Outcome
The system turns market noise into traceable candidates with visible reasons, stability, and limitations. Agents cannot push directly to main or execute wagers.
Market noise in. Reviewable evidence out.
Capture public market snapshots at a controlled frequency without republishing raw datasets.
Convert different market structures into a consistent model with timestamps and data-health states.
Combine movement, persistence, stability, and model divergence into candidates with explicit reasons.
A person chooses to watch, pass, or record a paper decision; the system performs no real transaction.
A focused review flow on mobile
The mobile view prioritises candidate reasons, timing, and human actions while reducing unnecessary table density.
Explainability before prediction
The interface shows why a candidate surfaced, how long the evidence has persisted, and where uncertainty remains. It is a review system rather than a promise of predictive accuracy.
Human authority stays explicit
AI agents may prepare isolated changes, run tests, and open pull requests. They cannot push directly to the main branch or perform any wagering action.
A deliberately synthetic public demo
Public screenshots and demonstrations use fictional teams and synthetic values. Runtime databases, account information, credentials, and raw operational records remain private.