Live dashboard: https://www.gildedacecasino.biz (backup: https://dashboard-royale.vercel.app). Click Judge / demo access.
On TAMU Wi-Fi? The campus network's DNS filter blocks newly registered domains, so the site may show as unsafe or fail to load. Use mobile data or an off-campus network.
Inspiration
A casino GM wanted to replace the 10 machines that earned the least. We asked whether that's bad machines, bad luck, or bad spots. Raw earnings can't tell those apart, so we built a dashboard that can.
What it does
- Machines: compares every machine with identical games in the same kind of spot on the same days, with luck removed. Seven of the GM's 10 just sit in a quiet zone, two were on the floor only part-time, and one (M-1158) leaks payouts.
- Demand, downtime, players, comps: for example, the May–June slowdown is fewer visitors (we ruled out six other causes), and a comp-price change on Apr 28 raised comp cost 12% with no change in visits.
- July plan: all 353 allowed mixes of priced actions, ranked on profit. We fund 2 promotion weekends + 1 inspection: $31k for about +$21k, with no loss even in the worst case.
How we built it
- Databricks medallion pipeline: raw Parquet → Unity Catalog bronze → silver → gold Delta tables, with shared SQL mirrored on DuckDB and parity-checked column by column (36/36 gold tables match).
- Models in MLflow: a GLM forecast frozen on Jan–Apr and tested on May–Jun against simple baselines, registered in Unity Catalog; leave-one-out models rank the machines; a LightGBM challenger is reported honestly.
- Statistics: expected earnings as ratios of sums, $\hat{r} = \sum \text{theo} / \sum \text{predicted}$, weekly t-tests with Holm correction, and shrinkage for thin data.
- Next.js on Vercel: staff sign-in, filters on every page, and every number linked to one of 31 claims with a re-runnable query or script.
Challenges
- Separating location from machine quality: a "bad" machine is often a good machine in a dead corner.
- Short periods over-flag machines, which we fixed with t-tests and same-day calibration.
- Being honest when our model only slightly beats a weekday average. We used it as the promotion baseline instead of overselling it.
What we learned
- Luck, exposure and availability change the answer more than model complexity.
- Natural experiments beat assumptions: a walkway closure showed that 69% of Back Forty's lost play just moved elsewhere, which killed our own signage idea.
What's next
Run the July plan, check each action against its pre-written success line, and replace the staff accounts with the casino's single sign-on.
Built With
- data-lake
- databricks
- duckdb
- lightgbm
- mlflow
- next.js
- numpy
- pandas
- pyarrow
- python
- react
- recharts
- scikit-learn
- sql
- typescript
- unity-catalog
- vercel