Inspiration

Casinos process enormous volumes of financial transactions, making it increasingly difficult to identify sophisticated money laundering patterns hidden within seemingly ordinary gambling activity.

My inspiration came from real-world financial crime investigations and regulatory warnings. In September 2026, the Financial Action Task Force (FATF) published findings highlighting how criminals exploit gambling platforms through transaction structuring, multiple accounts, coordinated activity, and financial movements that do not involve genuine gambling.

I built CASINOGUARD to address this challenge: an AI-powered casino financial intelligence platform that helps uncover suspicious patterns, connect financial activity, and turn complex transaction data into actionable investigative insights.

What it does

CASINOGUARD is an AI-powered casino financial intelligence platform that analyzes casino transaction data to identify potentially suspicious financial behavior and uncover patterns associated with money laundering.

I developed six specialized detection algorithms:

  1. Suspicious Transaction Chains: Identifies connected transactions that form potentially suspicious financial trails.
  2. Transaction Structuring: Detects attempts to split large transactions into smaller amounts to avoid reporting thresholds.
  3. Rapid Cash-outs: Identifies unusual deposits followed by rapid withdrawals or cash-outs.
  4. Unusual Transaction Volumes: Detects abnormal transaction frequencies and amounts compared with expected activity.
  5. Circular Transactions: Identifies patterns in which money moves between accounts and eventually returns to its original source.
  6. Multiple-Account Activity: Detects potentially coordinated financial activity across multiple player accounts.

Beyond detection, I built an interactive investigation dashboard that allows users to explore suspicious activity, examine transaction patterns, and understand why particular activity was flagged.

I also integrated SerpApi for public-information enrichment, allowing investigators to retrieve additional context about relevant individuals and entities.

I use Databricks for data processing and analytical workflows, and Tableau for interactive data visualization.

My goal with CASINOGUARD is to move beyond isolated transaction alerts and provide a more connected view of casino financial activity, helping investigators identify patterns that might otherwise go unnoticed.

DASHBOARD ROYALE CHALLENGE !!! -ACCESS https://casinoguardroyal.netlify.app

Q1. Which machines are truly underperforming once you account for exposure, availability and luck?

M-1008, M-1230, M-1045, M-1227 and M-1082 are the genuinely underperforming machines. I compared machines per available hour, excluded any with fewer than 90 days on the floor, and ranked them against machines with the same configuration. All ten lowest-ranked machines are in Back Forty. Re-ranked within that zone, only these five stay in its weakest 20%. The other five (M-1014, M-1040, M-1048, M-1265, M-1221) are ordinary for Back Forty, so their weakness is location.

Luck explains none of them: their payout z-scores run from −0.06 to +2.23. The manager's list is driven by exposure and one possible leak:

  • M-5076 was on the floor for only 49 days, M-1118 for 55 and M-1120 for 132.
  • M-1158 held 3.04% against an expected 10.00% (z = −4.62, below theo in 6 of 6 months). That is a reason to inspect it, not proof of a fault.

Q2. What changed in demand across the floor over the six months, and why?

Total floor demand was roughly flat, at $1.34M per day in January and $1.32M in June. It shifted between zones: Center Pit rose 10.5% while Back Forty fell 10.1% and the High-Limit Room fell 11.2%. The main cause I found is the Main Entrance–Back Forty walkway closure from March 3 to April 6. During it, Back Forty coin-in fell 29.4% and its traffic fell 30.4%, 32.9 points worse than the rest of the floor. After reopening it stayed 6.7 points behind, while Center Pit gained 8.7% during and 11.3% after. That pattern fits players re-routing, but individual guest routes aren't recorded, so it isn't proven.

Q3. How much play was lost to downtime, and when?

Machines were down for 5,021.08 hours across 1,226 tickets, of which 4,934 hours came from 1,207 fault repairs. Valued at each machine's usual theoretical win per hour, that is up to $88,346.72 of play lost. That is an upper bound, because some guests would have moved to another machine.

Downtime peaked:

  • By month: in May (1,009 hours).
  • By weekday: from Friday to Sunday.
  • By hour: in the evening, 18:00 to midnight.
  • By zone: in Center Pit, which had the most tickets and the highest lost play.

Ridgeline machines needed 7.23 tickets each, against 3.02 for Northwind and 1.93 for Halcyon.

Q4. What did tracked players wager, get back and net, and how concentrated are those results?

Tracked (loyalty-card) players wagered $163,804,869.65, were paid back $148,219,409.05, and the casino netted $15,585,460.60. That is only about $100K above expected, so luck roughly cancelled out. 80,992 players lost, 15,504 won and 17,394 broke even.

Results are highly concentrated:

  • The top 5% of players by wagering placed 55.34% of all wagers.
  • The top 5% of losers account for 40.73% of all losses.
  • The Gini coefficient of wagering is 0.7536.

This covers carded play only, about 66% of all coin-in.

Q5. What did the comp program cost, and did anything about it change?

The program cost $190,669.30 (face value $682,749.50), and the three source tables agree. Monthly cost rose from 1.11% of carded theoretical win in January to 1.32% in June. On May 1, the points needed per comp dollar halved in every tier: Base from 15.0 to 7.0, Gold from 12.0 to 5.6 and Platinum from 9.0 to 4.2. Average monthly cost rose from $30,812.60 (January–April) to $33,709.45 (May–June). The data contains no policy record of the change, and I can't say whether it brought in more visits.

Q6. Which actions would you fund with $100,000, and how would you know they worked?

Spend $17,500 and keep $82,500:

Action Cost Breaks even if… How it's measured
Inspect 5 machines (M-1158, M-1054, M-1210, M-1010, M-1201) $6,000 37% of their payout shortfall is a real fault Paytable and meter checks, then hold vs the configured rate, on the forecast's quietest July days
New signs for Back Forty $7,500 They recover 38% of the zone's remaining gap Zone traffic and coin-in vs other zones (difference-in-differences)
Comp pilot with a control group $4,000 Doesn't pay back in 30 days Comp cost per dollar of theoretical win and visits, treatment vs control; informs a decision worth up to $34,762 a year

I didn't fund replacements: half the candidates are only weak because of where they stand. An extra technician would recover at most $4,089 against its $11,520 cost.

DATABRICKS CHALLENGE !!! - In the YouTube video watch 4:48 - 5:30 for databricks

My Experience Using Databricks in CASINOGUARD

When I first started working with Databricks, I honestly found it intimidating. The interface had so many features, notebooks, compute options, and data engineering concepts that I initially thought it would be difficult to manage. I was worried that one small mistake could disrupt the entire workflow. However, as I started using it for our CASINOGUARD AML project, that impression completely changed. What initially seemed complicated turned out to be one of the most useful tools in our data processing workflow.

CASINOGUARD was designed to identify suspicious financial activity within casino transactions. We focused on six major AML patterns: suspicious transaction chains, transaction structuring, rapid cash-outs, unusual transaction volumes, circular transactions, and multiple-account activity. Databricks played an important role in preparing, processing, and analyzing the transaction data that supported these detection workflows.

I used Databricks notebooks as the main environment for developing and testing our PySpark operations. This allowed me to explore the transaction data, execute code in smaller sections, inspect intermediate results, and refine the processing logic without having to run the entire workflow every time.

One of the key features we implemented was the Medallion Architecture, consisting of Bronze, Silver, and Gold layers. The Bronze layer preserved the original transaction records. The Silver layer cleaned, standardized, and transformed the data to make it consistent and suitable for analysis. The Gold layer contained structured, analysis-ready datasets that supported our AML detection logic and downstream reporting.

Using PySpark, I performed operations such as filtering, grouping, aggregations, joins, and window-based calculations. These were especially useful for examining transaction frequency, comparing customer activity, identifying repeated patterns, and connecting related transactions. The processed data helped support the different detection mechanisms within CASINOGUARD, making it easier to investigate potentially suspicious behavior rather than looking at individual transactions in isolation.

We also used Databricks Jobs to organize and execute the data processing workflow in a repeatable manner. This helped connect the different processing stages and made the pipeline easier to maintain and validate. Instead of treating data preparation as a one-time task, we built a structured process that could be executed again as new transaction data became available.

Ultimately, Databricks was not just an additional technology in our project. It was the data engineering foundation that helped turn raw casino transaction records into organized information for AML analysis. I went from being overwhelmed by the platform to understanding how its notebooks, PySpark capabilities, layered architecture, and job workflows could work together to support a real-world financial intelligence application. That was probably my biggest takeaway: a tool that initially looked dangerous and complicated became much easier once I understood how its individual features fit into the bigger picture.

Challenges I ran into

One of my biggest challenges was distinguishing genuinely suspicious financial behavior from normal casino activity. A large transaction does not automatically indicate money laundering, and suspicious patterns often become apparent only when multiple transactions are examined together.

I addressed this by developing six specialized detection algorithms, each targeting a different financial behavior rather than relying on a single detection rule.

Another challenge was working with large, interconnected transaction datasets. I had to organize raw financial records, identify meaningful relationships between transactions and accounts, and perform time-based and statistical analysis. I used Databricks to structure these analytical workflows and make the data easier to process and investigate.

I also faced the challenge of translating complex analytical outputs into something understandable and useful for investigators. Building the detection logic was only part of the problem; presenting the results clearly through interactive dashboards and a centralized interface was equally important.

Finally, developing the entire platform independently meant balancing data engineering, algorithm development, frontend implementation, and integration while ensuring that all the components worked together.

Accomplishments that I'm proud of

I'm proud of building CASINOGUARD independently, taking the project from an initial idea to a functional financial intelligence platform.

Some of my biggest accomplishments include:

  • Six specialized detection algorithms: Developed individual algorithms to identify six different types of potentially suspicious casino financial activity.
  • Databricks integration: Used Databricks as the core analytical engine to process transaction datasets, perform complex aggregations, analyze financial patterns, and generate meaningful analytical results.
  • End-to-end analytical workflow: Connected raw transaction data, Python-based detection logic, analytical outputs, and interactive visualizations into one cohesive system.
  • Interactive dashboards: Built visualizations that make complex financial patterns easier to explore, compare, and understand.
  • Investigation-focused interface: Developed a centralized CASINOGUARD interface that brings financial intelligence and detection results together.
  • Independent development: Designed, implemented, and integrated the different components of the platform on my own.

Most importantly, I'm proud that I was able to turn a complex real-world financial crime problem into a working technical solution that combines data engineering, financial analytics, and software development.

What I learned

Building CASINOGUARD independently taught me how complex financial crime detection becomes when working with large, interconnected datasets.

I learned that identifying suspicious financial activity requires more than looking for unusually large transactions. Timing, transaction frequency, account relationships, behavioral patterns, and financial context all play important roles in identifying potentially suspicious activity.

I also gained hands-on experience with Databricks, particularly in data processing, analytical workflows, aggregations, and time-series analysis. Working with real-world-style casino transaction data helped me understand how data engineering can support practical financial intelligence applications.

Developing the six detection algorithms strengthened my understanding of statistical analysis, rule-based systems, anomaly detection, and transaction relationship analysis.

Beyond the technical aspects, I learned how to connect multiple technologies into a cohesive application, transform analytical results into meaningful visualizations, and independently manage an end-to-end software development project.

Most importantly, I learned that effective financial intelligence is not just about detecting anomalies; it is about providing enough context to understand why an activity deserves further investigation.

What's next for CASINOGUARD

I envision expanding CASINOGUARD into a comprehensive, intelligent financial monitoring and investigation platform for the casino industry.

My future plans include:

  • Machine Learning-Based Detection: Incorporate machine learning models to identify complex and previously unseen suspicious transaction patterns beyond predefined rules.
  • Real-Time Monitoring: Extend the current analytical workflows to support continuous transaction monitoring and immediate suspicious activity alerts.
  • Advanced Transaction Networks: Develop interactive transaction relationship graphs to uncover complex money movement patterns and connections between multiple accounts.
  • Explainable AI: Introduce AI-generated investigation summaries that explain why transactions were flagged, highlight supporting evidence, and help investigators understand suspicious activity.
  • Adaptive Risk Scoring: Develop dynamic risk scores that consider transaction history, behavioral patterns, account relationships, and multiple detection signals.
  • Expanded Data Integration: Connect additional financial data sources to provide investigators with a more comprehensive view of suspicious activity.
  • Improved Investigator Workflows: Build tools for tracking investigations, documenting findings, and managing flagged cases.

My long-term goal is to evolve CASINOGUARD from a transaction analysis tool into a proactive casino financial intelligence platform that helps investigators uncover sophisticated money laundering patterns and make better-informed decisions.

Share this project:

Updates

Submission history