Economist who builds machine learning that survives an audit. I read a P&L, I build the pipeline that feeds it, and I write the control that blocks my own model when it does not comply.
Three years inside corporate finance — treasury, billing and FP&A — supporting operations across 15+ countries in the Americas. MSc Economics at Pontificia Universidad Javeriana: thesis filed August 2026, graduation expected November 2026.
Bogotá · GMT-5 · full overlap with US hours · open to remote roles
A credit decisioning system on real US public data — 1.96M SBA 7(a) loans and 62.4M HMDA applications. The model is not the point. The point is that it survives an audit, and that I ran the audit against myself first.
My first honest number was a deleted one. The signal check returned AUC
0.9461, which is not a credit model, it is a leak: TermInMonths is overwritten
when a loan is liquidated, so the field carried the outcome. Removing it drops
the ablation to 0.6621. Production sits at 0.7005, +0.0311 over an
interpretable WoE scorecard, with calibration error 0.0107.
Ten promotion gates, and one of them blocks my own model. The access model scores a disparate impact ratio of 0.7639 against a 0.80 threshold, so it is not promoted — and I did not move the threshold. Every threshold has its derivation written next to it, and two of the gates exist because the governance documents lied: the model card listed 7 gates of 8, omitting the only one the model fails, while the validation report printed PASS for that same gate in one section and "promotion blocked" in another.
Declining the riskiest 10% would have avoided $276.3M in charge-offs, 2.15x what declining at random achieves, and $942.1M of the period's realized loss was absorbed by the taxpayer through the guarantee. It also forgoes $1.99B in good lending volume — 7.2x the loss avoided — and both numbers ship in the same payload. A headline that shows only the numerator is not a headline.
A charge-off takes a median of 51 months to appear, so performance monitoring
on young cohorts is arithmetic, not measurement, and the project refuses to fake
it. What it monitors instead found something: the SBA changed the business_age
category scheme between FY2018 and FY2021, and today 84% of its values fall
into categories the model never saw. Serving maps them to "unknown", so the
model does not degrade — it loses a top-three variable entirely and keeps
answering with the same confidence.
On the causal side I published a non-identification, not an effect. The guarantee percentage is the one lever the SBA actually controls, and it is administratively determined: R² 0.9145 against processing-method × loan-size cells, so there is no overlap left to exploit. At the $150,000 statutory threshold, 83.1% of the ±$5k window sits at exactly $150,000 — the density is destroyed, and with it the regression discontinuity. Conditioning on size removes 38% of the raw gradient and leaves a residual whose sign is what adverse selection predicts. An estimator applied where its assumptions fail produces a number, not an estimate.
Out-of-time validation across the COVID shock · under a 2007-style regime the
same model falls to AUC 0.5456 and underestimates risk eightfold, which is a
deliverable and not a caveat · 13 architecture decision records · make reproduce asserts the published metrics are identical after a retrain · CI was
red for eight consecutive commits before I caught it, and the fix was a script
that reproduces CI locally before pushing.
Python LightGBM PyTorch DuckDB PySpark MLflow ONNX FastAPI Power BI
financial-inclusion-colombia · open the atlas
I asked whether financial inclusion predicts departmental growth in Colombia.
It does not — β = 0.0007, p = 0.90, with entity and time fixed effects across 33 departments, 2019 to 2025, N = 228. Wild cluster bootstrap p = 0.89, permutation placebo p = 0.68. Without time effects the same coefficient is +0.024 with p < 0.001, and that distance is exactly what the national trend was worth.
I published the result with its specification, its N, its clusters and its tests. I also published what went wrong: sampling adequacy came to KMO 0.314 for access and 0.404 for use, below the 0.5 a factor model needs, so the planned PCA was dropped. Forcing it produced negative implicit weights on microcredit — an index that says more credit is less inclusion.
Nineteen public sources with a sha256 manifest · every series resolved to municipal codes, 100% coverage across the 34 quarterly cuts · an index by dimension with frozen, published weights · two annual panels · an atlas of all 1,123 municipalities · a full battery against spurious correlation (CIPS, CCE, permutation placebo, shift-share, event study, Moran, SAR/SDM).
Python dbt DuckDB linearmodels Quarto BigQuery
public demo, no account · private repository
A full day — habits, body, sleep, food, spending — logged in under ninety seconds with one hand, and handed back read rather than raw.
Half the people who start a tracking app abandon it in the first month, so the design targets specific failure modes: a gap is not a failure, a mastered habit graduates instead of counting as churn, nothing is interpolated, and no progress bar points at a target weight. A personal-finance data model plus health tracking on Postgres with RLS: 35 tables, 22 views, ~370 tests and an RLS smoke test in CI.
Next.js TypeScript Supabase RLS PWA
market-data-medallion · open the lab
Public APIs into a PostgreSQL medallion warehouse with dbt, an honest backtesting engine on top, and a daily refresh on GitHub Actions that keeps itself alive without a server. 48 assets, 58,000+ daily candles, zero budget.
Of 1,392 strategy variants evaluated, barely one in eight of the in-sample winners survived out-of-sample validation. I published every one that did not. The 42 variants combining five signals at once won zero times: more degrees of freedom is not more signal, it is more room to fit noise.
A signal computed at day t's close executes at day t+1's open — never at the close that produced it, and a regression test asserts that truncating the future does not change past signals. Fees and slippage are always on, and buy-and-hold pays the same. 89 dbt data-quality tests and 171 Python unit tests run before a single figure is published.
Python dbt PostgreSQL pandera GitHub Actions Power BI
- I publish what did not work. The null result, the KMO below the threshold, the data leak I found in my own app. The engineering log records 28 defects found and fixed, numbered one by one, and the credit-risk repository keeps its own ledger — including two published numbers I had to retract because they did not replicate on a second machine, and one claim I repeated four times before measuring it and finding it false. A portfolio that only shows wins says nothing.
- Every figure traces to a test. If a number appears in a document, it comes from a test, a row of the verification ledger, or a dbt test. In the credit-risk repository the gate does not trust the artifact either: it recomputes the published metrics from the saved predictions before letting anything be promoted.
- Decisions are written before the code. Sixteen decision records in the research repository and thirteen in the credit-risk one, each with the assumption that kills it if it fails.
- A control that cannot fail is not a control. Three of the defects I found in my own tooling reported success while doing nothing: a compliance check that matched nothing, a hook that accepted what it was built to reject, and a pipeline that printed "approved" after the training step had crashed. Each one now has a test that fails in the environment where it silently passed.
Remote Finance Data Analyst, Analytics Engineer and FP&A automation roles — where financial judgement and data engineering are paid as one capability, not two halves. Consulting too. I answer in English and Spanish.