Runtime
Local firstThe development workflow runs without hosted dependencies.Local-first NBA prop research
PRISM
A daily scanner that pairs quantile estimates with local box-score history and a clear record of settled outcomes.
- Role
- Modeling · Local data · Product engineering
- Year
- 2025
- Repository
- Private
01 / Interactive system slice
Inspect the workflow
Public sandbox · real engine with clearly labeled cached and preview states◇PRISMNBA prop scanner
LIVE TOOL · PUBLIC SANDBOXDETERMINISTIC PREVIEW
Starting the free-tier service…
Deterministic preview · estimates are distinct from settled outcomes
MIN EDGE (× MAE)1.00
3Evaluated
0Actionable
0Full value
TOP PLAYS
02 / Engineering evidence
Claims with boundaries
Derived from the local implementation and offline artifacts.Estimate
Quantile intervalLow, median, and high estimates remain visible together.Review
Picks → recordPredictions stay distinct from subsequently settled results.Problem
Why this exists
A ranked list can make model estimates look more certain than they are and blur the line between a prediction and a settled result.System
How it is bounded
SQLite-backed history feeds quantile models and a local API. The interface preserves prediction intervals, source context, and outcome status across the daily workflow.Outcome
What the demo proves
A compact research loop for scanning candidates, inspecting player context, and auditing what actually settled.Implementation surface
PythonLightGBMFastAPIReactSQLite
Building reliable data products?