Banking, research, consulting: the same question every time.

Fraud review kept re-deciding the same ambiguous transaction shapes because nobody had ever written down what made them ambiguous. I worked through 4.2 million transactions in SQL Server and Python and pulled out 11 recurring anomaly patterns that reviewers were encountering over and over, then turned them into a rulebook they could actually use at the desk rather than a deck they would read once. Manual review time dropped 30 percent, around 15 hours a week the team stopped spending on questions it had already answered.

Regulatory reporting runs on a schedule that does not move, and the parts of it still being assembled by hand each cycle are exactly where filing errors come from. I took ownership of six filings per quarter and rebuilt the Alteryx ETL workflow behind them with validation checks at each stage, so a bad value gets caught in the pipeline rather than in a submission. Those checks caught more than 40 data errors before they left the building, and the filings held zero reportable findings across four audit cycles.

Customers were abandoning mobile check deposit partway through and nobody could say at which step. I modelled the funnel on PostgreSQL data and built a Tableau dashboard that let the team walk the drop-off stage by stage instead of arguing from anecdotes, which put the blame squarely on ID verification. I took that to nine product and operations stakeholders with a specific fix rather than a finding. Completion rose 12 percent, roughly $50K a month in revenue that had been leaking out of one unfinished screen.

Twelve executive dashboards were being fed by 18 Azure pipelines with nothing between them, which meant a schema change or a short load became a number in a leadership review before anyone noticed. I added automated data-quality checks across those pipelines to catch schema and volume issues at the source, upstream of everything the dashboards do. Downstream reporting errors fell 40 percent, and the failure mode changed from a wrong figure in a meeting to an alert nobody outside the data team ever saw.

The roadmap debate hinged on an assumption nobody had tested: that customers who deposit through the mobile app churn at a higher rate than customers who come into a branch. I tested it against Snowflake data as a hypothesis rather than a hunch, then presented the result to seven senior stakeholders in a form that could withstand being challenged. That mattered, because it was going to move money. The finding supported a $2M product roadmap decision, and satisfaction scores rose 18 points off the back of what it funded.

A literature review of 200-plus papers is unusable as prose. You cannot count anything in it, and reading it sequentially takes about six weeks that the research calendar does not have. I built a Python pipeline that ran LLM summarization, keyword mining, and thematic clustering across the whole corpus, which turned the review into something you could query rather than something you had to finish. It came down to five days, and more importantly it surfaced cross-cutting themes in forecasting, supplier risk, and automation that nobody reading in order would have connected.

The nightly batch took six hours, which meant it finished somewhere around the time the client's analysts wanted to be using it, and when it broke there was no room left in the window to rerun. I rebuilt all 12 ETL jobs in Alteryx and Python, restructuring the work so independent stages stopped waiting on each other and a failure in one source did not take the run down with it. It finishes in 45 minutes now, holds 99.5 percent uptime, and removed about 10 hours a week of manual data preparation that existed only to patch what the pipeline had missed.