15 articles
An interactive comparison of a Poisson GLM and XGBoost on 678,013 motor policies: better risk ranking, SHAP explanations, a geographic fairness check, and the regulatory limit that keeps ML a pricing challenger.
Traditional credit analysis treats each loan as independent, but guarantee chains, circular guarantees, and ownership concentration create correlated exposure that relational models cannot express. This project models a 500-client portfolio as a Neo4j knowledge graph, processed with PySpark on Databricks and scored with calibrated LightGBM, to surface structural risk patterns that SQL keeps hidden.
Interpreting LISF and CUSF means navigating articles that cross-reference each other across laws, and a Ctrl+F can't tell the article defining technical reserves from one that mentions them in passing. AI makes it possible to absorb that entire volume without losing a single detail. This agent uses RAG to index every article individually with a cross-reference graph, eliminating citation hallucinations and ensuring the model only reasons over real legal text. The result is an assistant that amplifies the actuary's memory without replacing their judgment.
How classifying 5.1M Major Medical Expenses claims into three hospitalization levels changes the way you price a risk the industry treats as one. A UNAM team project that became a complete pricing system.
In Mexico, roughly 70% of vehicles circulate without insurance. For the insurers covering the rest, the business boils down to three questions: how much to charge, how much to reserve, and where the fraud is. This platform builds a dashboard that answers all three with data calibrated to the Mexican market, separating frequency from severity for pricing, estimating what remains unpaid with two complementary methods, and flagging anomalous claims before they reach adjustment.
Financial risk models lose credibility when they exist only as formulas in a PDF. These 13 modules implement them in typed Python with automated tests: from portfolio VaR and Monte Carlo simulation to copulas, EVT, deep hedging, and graph neural networks for systemic contagion. Each module pairs LaTeX theory with reproducible results on public market data.
There is no open-source actuarial library built for Mexican regulation. suite_actuarial fills that gap: it covers eight insurance domains (life, P&C, health, pensions, reserves, reinsurance, regulatory, and configuration) with EMSSA-09 mortality tables, CNSF circulars, and SAT tax articles built into the design. It installs with pip, deploys with Docker, and exposes 28 REST endpoints alongside a bilingual Next.js dashboard.
Actuarial teams generate valuable data that gets trapped in spreadsheets and manual processes that do not scale. This platform builds the complete pipeline on GCP, from streaming claims intake to Tweedie GLM pricing, with Dataform and BigQuery as the backbone. The result is an automated, tested, and reproducible flow that turns raw data into inputs ready for the regulator.
Airlines generate millions of flight, delay, and revenue records, but analyzing that data requires choosing the right database for each question. This project takes 5.74M real records, analyzes them first in PostgreSQL with engine-level optimization, migrates to BigQuery to compare both paradigms, and presents the trade-offs with real timing, real costs, and real query plans.
Most Mexican workers do not know which pension regime they contribute under or what they will actually receive at retirement, and official sources do not simplify the comparison between Ley 73, Ley 97, and Fondo Bienestar. This simulator implements all three formulas with current data (UMA, CONSAR tables, EMSSA 2009 mortality) and lets users explore scenarios with interactive sensitivity analysis. The result is an educational estimate that shows what you control and what you do not.
SIMA centralizes actuarial techniques for pricing life insurance: it takes raw mortality data from INEGI/CONAPO, graduates it with methods like Whittaker-Henderson and Lee-Carter to obtain curves that respect human biology, and projects forward to calculate premiums, reserves, and capital requirements under LISF. Everything exposed as an API, allowing it to connect with other systems, automate sensitivity analysis, and meet CNSF requirements. Open source and built to expand into other lines of business.
Frequency-severity pricing models on freMTPL2: Poisson GLM vs XGBoost vs LightGBM with SHAP explainability, fairness audits, and a cross-border analysis of what European ML pricing techniques mean for Mexico's 70% uninsured auto market.
Seven end-to-end projects across six domains: real estate, insurance reserves, e-commerce cohorts, product experimentation, finance (SaaS KPIs and portfolio risk), and public-sector operations. The same statistical toolkit runs through all of them, and that is the argument: a data analyst has to hold several domains in view at once, because range is what makes depth transferable. Each project ships as an interactive dashboard, all on one live domain.
Actuarial reserve analysis using Chain-Ladder and Bornhuetter-Ferguson methods on NAIC Schedule P regulatory data. Interactive dashboard with loss triangles, IBNR estimates, and combined ratios across 6 lines of business.
I wanted to understand what really happens inside a language model. I built one from the first matrix multiplication, trained it on all 7 volumes of Proust, and what taught me the most wasn't the architecture; it was realizing that everything is just numbers.