My work spans behavioral economics, operational research, and product analytics. I've tested pricing psychology across 185,000+ transactions and quantified weather's impact on transit demand to save £23K/month.
See the workTechnical Skills
Projects
A record of questions I asked, methods I used and what I actually found.
Beyond the Treadmill: A Multivariate Analysis of Exercise and Mental Health
Does exercise actually help depression, or are we missing the bigger picture in the data?
Finding: Exercise is a significant predictor of depression (p = 0.014) but explains only 0.25% of variance in isolation. A whole-person model revealed Age and BMI are mathematically stronger predictors. Clustering exposed a High-Risk group that exercises maximally yet shows 67.5% clinical depression prevalence. Three categorical hypothesis tests failing in sequence is not a dead end - it is proof that continuous data must not be bucketed.
Impact: Demolished the simplistic "exercise equals happiness" narrative with real-world NHANES data (n = 2,004). The methodological finding - that forcing continuous data into categories destroys statistical signal - is the most transferable result.
Feature Redundancy in Breast Cancer Diagnostics
Clinical labs record 30 features per tumour sample. Are all 30 doing independent work, or is the dataset carrying a lot of redundant weight?
Finding: Radius, perimeter, and area are near-perfectly correlated (r ≥ 0.98) because they are geometrically linked. Measuring all three is statistically measuring one thing three times. PCA confirmed the dataset compresses to 10 components at 95% variance explained. Welch's t-tests ranked all 30 features by diagnostic power: worst concave points led by a wide margin (t = 29.12), while fractal dimension and symmetry failed Bonferroni correction entirely.
Impact: Redundancy in clinical measurement inflates perceived complexity and can obscure which features actually drive diagnosis. This analysis identifies exactly which features earn their place and which don't.
Weather Impact Analysis: Hypothesis Testing
How much does weather actually affect UK public transit demand?
Finding: Weather accounts for 38.7% of demand variance across 17,400+ hourly observations. The pattern is consistent and predictable enough to schedule around.
Impact: Turned a vague operational problem into a staffing model. Dynamic scheduling based on weather forecasts saved £23K/month. Rigorous assumption-checking is what made the result trustworthy enough to act on.
Left-Digit Bias Audit: Pricing Psychology
Does X.99 pricing actually work? I analyzed 185,000+ transactions to find out.
Finding: It works for some products, not others. Phones showed a +5.16% sales lift with X.99 pricing (p<0.001). Laptops showed no effect at all. Pricing psychology is category-specific, not a universal lever.
Impact: Retailers applying charm pricing uniformly are leaving money on the table in some categories and wasting it in others. This analysis tells you exactly which is which.
Bayesian A/B Testing: E-Commerce Optimization
Does reducing product density improve conversions? And how confident can I be in the answer?
Finding: 98% probability that lower product density improves conversion rates. The sensitivity analysis held across multiple priors, meaning the result is not fragile to starting assumptions.
Impact: A frequentist test tells you whether the result is significant. This tells you the probability that the decision is correct. This matters for real-time business decisions.
Retail Operations and Finance Optimization
Where are the bottlenecks in this $66.31M operation? Why are fulfillment times inconsistent?
Finding: Specific warehouses were responsible for 40% of delays. Not a systemic failure across the operation. One identifiable root cause, isolated through segmentation.
Impact: Reduced manual data processing by 60%. More importantly, the question changed from "why are things slow?" to "warehouse X has a process problem, here is what needs to change."
Education
BSc in Mathematics
University of Nairobi