I turn technical findings into decisions people can act on — starting with what a decision-maker needs to know, confirming the data can actually answer it, and finishing with a recommendation.
In recent projects I found that a retailer's largest customer segment wasn't its most valuable one, and that a scholarship programme reliably drove enrolment while having no measurable effect on student progression — the kind of finding that changes where money goes.
What I bring
A decision-first approach that keeps the choice someone needs to make in view.
Python and Excel to clean and prepare data, with working SQL knowledge for querying.
Regression, hypothesis testing, and segmentation in Python and R to separate signal from noise.
Interactive Tableau dashboards that make findings clear to stakeholders.
Comfortable presenting findings and defending the reasoning behind them.
Most interested in work where data shapes future strategy rather than only reporting on the past. Based in Nepal, UTC+5:45 — Asia-Pacific in my morning, Europe through my afternoon, the US East Coast in my evening.
Determine whether scholarship discounts actually drive student enrolment and progression, so allocation decisions rest on evidence rather than assumption.
Approach
EDA on historical application data, Chi-Square tests of independence for enrolment and progression outcomes, OLS regression, and regional/area-level conversion analysis.
Result
Students offered a scholarship enrolled at 70.5% versus 6.2% without one — an 11× difference against a 9.99% baseline. But the same testing found no statistically significant effect on academic progression: scholarship money buys enrolment, not student success. Effectiveness also varied sharply by market — China converted strongly on low discounts, while South Asia and UK, EU & Americas underperformed relative to spend.
Recommendation
Allocate scholarships by historical conversion rate per region rather than a uniform global model.
Help the marketing team design a campaign by identifying who their customers are, what sells, and which advertising channels convert.
Approach
Cleaned and validated 2,213 customer records in Excel (repairing malformed date and currency fields, removing implausible ages), loaded to PostgreSQL for channel analysis, and built two interactive Tableau dashboards with demographic filters.
Result
Across $1.34M in sales, alcoholic beverages and meat products drove 78% of revenue (50.3% and 27.5%). Spain is the largest market — 49% of customers, and the biggest share of sales of any country. The key finding: sales climb with income only as far as the $70–79k band and decline above it, so the highest earners are not the most valuable customers. Twitter was the most effective ad channel; brochures the least.
Recommendation
Concentrate spend on Twitter targeting married 50–59-year-olds in Spain, and prioritise the $70–80k income band over both the largest segment and the highest earners.
Identify the customer trends driving loyalty point accumulation, and gauge sentiment toward the brand, to improve overall sales performance.
Approach
Correlation analysis across age, remuneration and spending score; distribution testing (skewness/kurtosis) to check normality assumptions; k-means segmentation on remuneration and spending; NLP sentiment analysis on customer summaries and reviews using TextBlob.
Result
Spending score (r = 0.67) and remuneration (r = 0.62) were the strongest predictors of loyalty points, while age showed no meaningful relationship. High earners accumulated 3.5× the loyalty points of low earners (2,503 vs 725 on average). Segmentation surfaced five distinct customer clusters, and sentiment averaged about +0.22 across both reviews and summaries — favourable but lukewarm rather than enthusiastic.
Recommendation
Introduce tiered redemption (5–20% discounts) so customers see tangible return on points, converting loyalty accumulation into repeat spend.
It began as a claim checker for NEPSE-listed equities, every figure carrying the sentence it came from. It ended as a study of a narrower and more awkward question: how do you know your metric measures what you think it measures? Submitted to the micro1 Frontier Engineering Challenge 2026.
Approach
Six checks, each built because the previous one returned “you do not actually know that”: a baseline with no documents at all, a decontamination of the source files, a control with the correction loop disabled, 23 tests over the validator, an audit comparing the validator’s two callers, and a dated record of the one time the ground truth moved. 384 lines of agent; 548 of apparatus.
Result
All six came back against me. The metric was at its ceiling before I started — a document-free baseline scored full marks, 3/3, and verdict agreement then read 3/3, 2/3, 3/3, 3/3, so nothing I built afterwards could show up in it. The correction loop cost 3.4× the time and moved nothing scored. And the checker driving that loop turned out to be stricter than the scorer, in four of eight attempts and never the other way.
What it taught me
That the findings are local and the method is what transfers. Three cases and one model support no general claim about agents — but the discipline does: ablate the part you believe in, keep a control arm, and point the same scepticism at the instrument doing the measuring. Every figure here is reproducible from the repository without an API key.
Tools: Python · Git · unit testing · experiment design · LLM API
Most interested in work where data shapes future strategy rather than only reporting on the past. Based in Nepal, UTC+5:45 — Asia-Pacific in my morning, Europe through my afternoon, the US East Coast in my evening.