Senior AI/ML Engineer at Nequi (Grupo Bancolombia), pre-training foundation models on customer transaction sequences. 8+ years shipping production ML. Based in Colombia 🇨🇴.
I'm most interested in the gap between what a model appears to do and what it actually does — evaluation that survives scrutiny, agent tool-use that earns its place, and systems whose failure modes are written down before someone finds them.
RealH — Proof-of-personhood and
content provenance using W3C Verifiable Credentials. Ed25519 signing, did:web,
public JWKS for offline verification. The README opens with four specific reasons
not to trust it in production, because a reference implementation that hides its
threat model is worse than none.
TypeScript · CI · CodeQL · gitleaks · ADR on JCS canonicalization
Kaggle write-ups — 19 competitions, each documented as a short paper. Includes a provenance ledger that states, per competition, exactly which submissions were my own work and which were not — verified by diffing my final submissions against the public notebooks they came from. 5 of 19 were substantially original; the ledger says which, and names every upstream author.
EML translator —
Library for translating mathematical expressions to and from Exp-Minus-Log trees
(Odrzywołek's primitive). Forward/inverse
translation, complex-branch evaluation, exhaustive identity search.
Python · 116 tests · MIT
EML research POC — Wires those operations into Claude as tool-use and benchmarks whether it helps. The honest answer, reported in the README: on calculus and algebra the agents largely ignore the EML tools in favour of sympy. It earns its keep only on translate/evaluate/verify tasks.
barrio-mapper — Maps the
small independently-owned businesses of a neighbourhood — tiendas, droguerías,
ferreterías — while filtering out chains. Three-signal chain detection with an
auditable rejection log. Tested in Medellín;
runs in the browser.
Python · tests · CLI + browser app
Production ML — foundation model pre-training on transaction sequences · credit risk (PD/LGD, loss-rate forecasting, IV/WoE, Platt calibration, SHAP-driven adverse-action reasons) · recommender systems
Agentic systems — tool-use design and evaluation · structured output · multi-step orchestration · guardrailed tool access · agent security (multi-step tool attacks)
Evaluation & observability — offline eval sets wired into CI · faithfulness and LLM-as-judge · embedding drift detection · PSI monitoring · Langfuse tracing
Stack — Python · SQL · PyTorch · scikit-learn · XGBoost · LightGBM · Hugging Face · BigQuery · dbt · Airflow · FastAPI · MLflow · AWS · Docker · Terraform · GitHub Actions
Portfolio · LinkedIn · Kaggle · TEDx talk
Trilingual (ES/EN/PT) · Mentored 13+ engineers