Nalin
Verma

Click A Diet↗

I launched an AI service that emails subscribers a personalised 28-day diet plan every month, now reaching around 200 paying customers. Quality was the hard part. I built an evaluation harness that pairs each generated plan with a blind second-model recheck, lifting the pass rate from 5% to 89%. A verification layer sits in front of the customer and blocks any plan failing a calorie floor, a macro target or a condition-specific limit, which cut nutritional error by roughly 95%.

Tiger Analytics

I turned transaction-level records into model-ready customer features, writing SQL and Python pipelines that leaned on window functions and CTEs over large banking datasets. That work fed a client's personalized-offer project: I assembled the customer dataset behind the logic deciding which offer each customer saw, and through which channel. Every deliverable shipped with its methodology, assumptions and findings documented, so a reviewer could trace the path from raw data to final output.

WAT.ai↗

I built the team's baseline regression model in scikit-learn, mapping 8 macroeconomic indicators onto 6 market targets across 10 years of monthly data. The more useful result was a negative one: I found CBOE Volatility Index data leaking into the equity targets, which had inflated our test R² to +0.30. Removing it revealed the true value of −0.52, and that honest baseline is what the rest of the work now builds on.

Purple MicroPort Cardiovascular

My first time using data to drive operational decisions in a regulated industry. I wrote Excel VBA scripts that automated budgeting and cost-tracking during finance reviews, cutting about 20% off the monthly reporting cycle, then broke spend down across cost centres so the finance and operations team could see where costs were actually concentrated.