The problem: A normal forecast gives one number for the end of the month. Real life does not: salaries vary, spending varies, and surprise bills arrive. One number hides the risk of running short.

What was done: Simulates the coming months thousands of times in TypeScript, with income and spending that vary and surprise expenses that strike at random. It can learn those patterns from an uploaded bank statement — including public Kaggle personal-finance datasets.

What changed: Shows the range you are likely to land in and the chance of dipping below a safety buffer, instead of a single guess. Runs in the browser on illustrative or uploaded data; it is a model, not financial advice.

The problem: Money laundering hides inside ordinary-looking transactions. Reviewing them by hand does not scale, and the patterns that matter are easy to miss.

What was done: Building a detection model in Python (Pandas) that works on synthetic banking data. It applies Benford's Law to the digits of transaction amounts, plus anomaly detection, to rank transactions by how unusual they are.

What changed: Produces a shortlist of high-risk transactions for a human to review, instead of asking someone to read every row. Tested on synthetic data only — no real customer data is used.

The problem: Invoice processing was slow and inconsistent, holding up payments to suppliers and tying up staff time.

What was done: Led a simulated process improvement team through a Lean Six Sigma transformation, using Minitab to run hypothesis tests and confirm that the changes made a real statistical difference rather than a lucky one.

What changed: Invoice processing time reduced by 30% in the study, with a projected annual saving of £50,000. These are capstone figures from a controlled exercise, not a live client result.

The problem: Reconciling thousands of transactions a month by hand is slow, repetitive, and the place where mistakes quietly creep in.

What was done: Wrote a Python script that reconciles more than 5,000 transactions a month automatically, and built an exception report that flags only the items that do not match.

What changed: Saves around 15 hours of manual work a month, and every unmatched item is surfaced rather than missed — so the audit trail stays complete.