5,821 data points sampled from my photo. Hover or touch them.
I build risk models, take them to production, and design the tools that let business teams use them on their own.
Ask my AI about my work
Runs on a Cloudflare Worker and answers only from information Kelly verified.
Projects that made it to production
Real work, real results. Each case names where it happened. Visuals are recreated with synthetic data.
Pacífico EPS, 2025 to present
Risk scoring the business runs without code
No-codemodels run by business teams
Read the case
Problem
Running a risk model required the data team every time, which slowed adoption.
Approach
Built risk scorecards and logistic regression models (WoE binning, validation with KS, Gini and ROC) and deployed them in production behind an interface built with Flet in Python.
Result
Business teams score cases on their own, and each variable’s contribution is readable by non-technical users.
Stack
Pythonstatsmodelsscikit-learnFletSQL
Pacífico EPS, 2025 to present
From two weeks to two hours
2 wk → 2 hplus 9 colleagues trained
Read the case
Problem
A recurring process took two weeks of manual work.
Approach
Rebuilt it as a Python pipeline, then ran Python training sessions for 9 non-technical colleagues so the team could own it.
Result
The process now runs in about two hours, and the knowledge no longer depends on one person.
Stack
PythonpandasOracleSQL Server
Logistics (MDPI), first author, March 2026
Random Forest + Genetic Algorithms for inventory
+4.52 ppROI vs. baseline, 5 of 5 stores
Read the case
Problem
Can machine learning beat a simple moving-average replenishment rule in real convenience stores in Lima?
Approach
A Random Forest forecasts daily transactions; forecasts feed an inventory simulation where a Genetic Algorithm tunes the monthly replenishment policy under coverage constraints.
Result
Forecast MAPE under 13%. Beat the 28-day moving-average baseline on ROI in all five stores, significant under the Wilcoxon test.
Take generative AI from experiments into the platforms teams use every day.
Approach
As the team’s AI Champion, developed and integrated use cases through the ChatGPT Enterprise API into real-time platforms.
Result
AI use cases running in production tools, and two first places in Latin America-wide internal hackathons.
Stack
ChatGPT Enterprise APIPythonPrompt design
Belcorp, 2024 to 2025
Campaign data from three ad platforms, automated
3 APIs → 1automated ingestion across LATAM
Read the case
Problem
Digital campaign data for several LATAM markets was collected by hand.
Approach
Integrated the Meta, TikTok and Google Ads APIs with Python into a single ingestion flow.
Result
Manual data collection eliminated across LATAM markets.
Stack
PythonMeta APITikTok APIGoogle Ads API
Pacífico EPS, 2025 to present
Interactive reports for management
1 reportinstead of scattered sources
Read the case
Problem
Management decisions needed data spread across several sources.
Approach
Built automated HTML reports that pull those sources together into one interactive view.
Result
Management reviews one up-to-date report instead of assembling the picture by hand.
Stack
PythonHTMLSQL
Lab: a credit scorecard you can play with
The same kind of model I build at work, rebuilt with synthetic data. Change the applicant and watch the score, the probability of default and the explanation update.
Synthetic bins and weights, built for this demo. No real customer or company data.
Score
604
Probability of default
1.8%
Medium risk: manual review
449682
What moves this score, compared with an average applicant
Age32
Debt-to-income45
Time at current job35
Late payments, last 12 months70
Credit line usage42
Raises riskLowers risk
How it works: each variable is binned, and each bin carries points derived from its Weight of Evidence and the logistic regression coefficient. Points add up to a score scaled so 600 means odds of 50 to 1, and every 40 points doubles the odds (PDO 40). Probability of default = 1 / (1 + odds).
My profile, explained like a model
A SHAP-style waterfall. The weights are symbolic; the evidence behind each bar is real. Select a bar to see it.
Starting point
0.50
Prediction
0.91
Models the business runs on its own
Risk scorecards and logistic regression models deployed behind a Flet interface, so business teams score cases without writing code.
Pacífico EPS, 2025 to present
Teaching and social impact
9
Colleagues trained in Python
Non-technical teammates at Pacífico EPS, trained to bring Python into their work.
7
Students in Pequeños Solucionadores
A pilot I designed and led on Design Thinking with AI for students from vulnerable communities.
EnFoco
Free platform for parents of children with ADHD
Co-leading it with a clinical neurologist and a fellow engineer, since February 2026.
PUCP
Teaching Assistant
Guided student teams through market studies, with voluntary advising hours.
Graduation day at PUCP, first in my class.
Path so far
Dec 2025 to now
Data Scientist, Pricing & Analytics, Pacífico EPS
Risk scorecards, statistical validation of A/B/C tests, automated reports for management.
Aug to Dec 2025
Teaching Assistant, PUCP
Market studies with student teams.
Jul 2024 to Dec 2025
Omnimedia Analyst, Belcorp
AI Champion. Generative AI use cases, ad platform APIs, and support on the Marketing Mix Model with Deloitte.
Aug 2023 to Jun 2024
Internal Audit Intern, Belcorp
An R script built from scratch that cut audit execution time in half.
2019 to 2024
B.Sc. Industrial Engineering, PUCP
Ranked 1st in the 2024-1 class. Beca 18 scholar. Aspire Leaders Program.
Let’s build something
Open to data science and AI roles, research collaborations, and graduate programs.
