Writing
5 posts
A synthetic loan book gave me 0.892 AUC. Real applicants gave me 0.665. Here is why the lower number is the one in production, why the setting that hit 88% accuracy was the worse product, and what happened when I tested whether any of it generalized.
Banks already have all your transaction data. The problem isn't capture - it's that the data sits behind clunky interfaces nobody opens twice a month. FinanceAI is what happens when you treat that data as a product instead of a statement.
I study data engineering. I work in sales. Most people treat those as opposites. Here's why that combination is the most useful thing on my resume - and what it taught me about why most data work fails before it starts.
Predicting business performance sounds straightforward until the model is right and the business still makes the wrong call. Here's what building a predictive financial tool taught me about the gap between data and decisions.
Bad data costs the global economy trillions every year. Building a tool to fix it taught me more about real-world data engineering than any course ever could.