Write-Up
Inspiration
The Gilded Ace, a fictional Las Vegas casino, had $100,000 to invest and ten underperforming slot machines. Replacing them seemed obvious, but we wondered: what if the machines weren't the problem?
Raw earnings don't account for location, operating hours, or payout luck. We built ante to look beyond surface-level numbers and make better investment decisions.
What it does
ante is an interactive casino decision-intelligence platform that analyzes machine performance, foot traffic, downtime, player behavior, loyalty costs, and investment opportunities.
Our biggest finding: none of the original ten lowest-earning machines remained in the demand-adjusted bottom ten.
Instead of replacing them, ante recommends inspecting one machine and relocating three. The plan spends $6,600 of a $100,000 budget, with a central projected net theoretical-win improvement of $13,272 and a scenario range of $2,059–$22,724.
How we built it
We developed a Databricks Bronze–Silver–Gold pipeline, ingesting 31 datasets, validating 31 Silver tables with 222 passing checks, and producing 12 Gold analytical tables.
Using Python, Streamlit, and Plotly, we built seven interactive analytical pages powered by a validated Gold snapshot.
We integrated the Google Gemini API to create Ask ante, a floating AI analyst that answers questions using verified evidence, source references, and numerical safeguards.
Challenges we ran into
Our biggest challenge was separating genuine underperformance from misleading statistics. We accounted for machine availability, changing locations, payout variance, and differences in customer demand.
We also had to distinguish correlation from causation, prevent double-counting in financial calculations, and ensure Gemini never presented hypothetical returns as guaranteed earnings.
Accomplishments that we're proud of
- Built and deployed a complete, interactive dashboard with seven analytical pages.
- Created a validated Databricks pipeline with automated reconciliation checks.
- Integrated a Gemini-powered assistant grounded in our analysis.
- Identified a targeted investment strategy that preserves 93.4% of the budget.
- Made our forecasts, uncertainty, and rejected alternatives transparent.
What we learned
More complicated doesn't always mean better.
Our ridge regression model performed better during training-only validation, but a simple weekday-average baseline achieved lower error on the untouched holdout: $68,699 versus $73,652 MAE.
We learned that strong data science means questioning assumptions, validating results, and recognizing when a simpler solution works better.
What's next for ante
We want to measure the recommended relocations over 30 days, compare results against similar machines, and account for displaced demand. Those outcomes would inform whether the casino should expand the strategy or keep its remaining budget unspent.
