Davirson Novoa Ramírez

$276.3M in avoidable charge-offs · AUC 0.7005 · 10 gates, and one blocks my own model

Credit decisioning system with model governance — design and build

• Declining the riskiest 10% of the test portfolio would have avoided $276.3M in charge-offs — 2.15x what random declines achieve — and the report ships the counterweight too: $1.99B in good lending forgone.
• Deployed loss model on 1.96M SBA 7(a) loans: AUC 0.7005 out of time, +0.0311 over an interpretable WoE scorecard, with calibration error 0.0107. The second model, for credit access over 62.4M HMDA applications, is the one the fairness gate refuses to promote; the event study runs separately, over 93.4M applications.
• The first reading was 0.9461, and it was not a model but a leak: TermInMonths is overwritten when a loan is liquidated, so the field carried the outcome. Removing it drops the ablation to 0.6621 — and that is the number published.
• Ten promotion gates judged by exit code, one of them designed to block my own model when it misses the threshold.
• Monitoring caught the source changing its vocabulary without notice: a field stopped meaning the same thing between vintages, and nobody would have known without the control.
• The production ONNX artifact scores in the reader's own browser — no server, and no data leaves the page.

1,392 strategy variants validated out of sample · 48 assets a day · 89 quality tests before anything ships

Market data platform — designed, built and operated

• Watermark-based incremental ingestion from 4 sources (Tiingo, Coinbase, Kraken, FX) into a three-layer PostgreSQL warehouse: 48 assets and more than 58,000 candles, idempotent and resumable.
• 89 automated dbt data-quality tests plus 171 Python unit tests; the pipeline fails before publishing bad data, not after.
• No-look-ahead backtesting engine with 70/30 out-of-sample validation over 1,392 strategy variants: the vast majority of in-sample winners did not survive the blind window.
• FX decomposition for Latin American ADRs, separating company performance from the currency move through the identity (1+r_USD) × (1+r_FX) = (1+r_local).
• Daily orchestration on GitHub Actions with a rate-limit circuit breaker, running on zero-cost infrastructure.

2,741 tests across 8 CI gates · 65 tables and 46 views under row-level security · a public demo that never touches the database

Own product — designed, built and operated

• Multi-user model on Postgres with row-level security on all 65 tables: the conversion required rebuilding 22 primary keys that identified rows without the user, and pushing the user inside the foreign keys so isolation did not rest on the policy alone.
• Eight CI gates judged by exit code: types, format, lint, 2,741 tests, build, dependency audit, a guard that hunts personal data inside the built artifact, and a smoke test requiring the anonymous role to be denied across all 111 tables and views.
• A public demo running the five real screens over 120 seed-generated days that never queries the database; a test walks the runtime import graph and fails if any reachable module gets to the data client.
• Day logging in one form and one button, with eight independent writes, an on-device draft and send queue to survive losing signal, and a purpose-built number reader because a browser number field was storing expenses divided by a thousand.

19 public sources · 1,123 municipalities · 160 specifications, each with its power

M.Sc. in Economics thesis, Javeriana — reproducible research end to end

• Purpose-built open-data downloader with a verifiable manifest: nineteen sources (financial supervisor, statistics office, ICT and education ministries, national map), row counts verified against the source, metadata-driven typing and a sha256 per file.
• Dimensional warehouse in dbt on DuckDB: department, municipality and period dimensions, financial inclusion in long form by product block, and tests that reconcile each department total with the sum of its municipalities.
• Geographic keys resolved to the municipal code with 100% coverage across the 34 quarterly cuts, and the splice between the supervisor's two tables measured quarter by quarter.
• A financial-inclusion index by dimension and annual panels at department and municipality level, with an explicit battery against spurious correlation. Public logbook of mistakes and wins.