Daniel Felipe Sacristán Ávila — Helicopter Pilot to Data Scientist

Streamlit dashboard integrating three independent data sources, with SHAP-explained models and stationarity testing (ADF/KPSS). Sleep-quality classifier scores ROC-AUC 0.77 and 0.70 accuracy under 5-fold cross-validation.

That AUC swings from 0.56 to 0.92 across folds (±0.15) on a small personal sample — the model is directional, not diagnostic. Same discipline applies to the data: no forward-filling, no assumed "today".

🧪 Nothing to upload. One click loads a fully synthetic dataset — every tab, model and SHAP plot is explorable with zero real health data.

BETO fine-tuned on 2,935 fact-checks harvested from ColombiaCheck's ClaimReview markup, published on the Hugging Face Hub with model card, datasheet and data statement.

macro-F1 0.405 vs 0.386 for the TF-IDF baseline — but "Verdadero" scores 0.00 with only 93 examples. It is a Falso/Cuestionable discriminator, not a truth detector, and the model card says so.

Self-updating GitHub Actions pipeline that backfilled 1,107,336 raw records into 101,256 continuous hourly observations of Colombia's spot price (precio de bolsa) since 2015 — no gaps, no nulls — and refreshes daily with no server.

XM's settlement versions don't mature in name order: TX3+ adjustments are published after the TXF invoice, so the naive ordering left 47% of hours stale.

MSc thesis: NLP over NASA ASRS incident narratives for taxonomic coverage of rotorcraft operations.

Evaluation is disaggregated by slice and audited against the ~250 reports covering aircraft I have flown — ASRS is voluntary and self-selected, so it cannot measure rates.