Data science · Statistical modeling · Auditable GenAI
I build statistical models, forecasts, interpretable machine learning and auditable Generative AI workflows—with fair baselines, honest uncertainty and reproducible evidence.
My default: make assumptions visible, outputs testable, and decisions explainable.
01 / Personal projects
Each case study starts with a decision, documents its constraints, and ends with reproducible evidence.
A leakage-aware propensity model that ranks outreach under a finite contact budget—evaluated with calibration, lift, capture, uncertainty, and transparent value assumptions.
Price elasticity on one million real retail transactions, a volume-discount bias found and corrected, and a price test sized with CUPED—45% fewer customers for the same power.
14-day national demand forecasts: SARIMAX with Colombian holidays, LightGBM and an ensemble, compared with Diebold-Mariano tests over a 2-year rolling backtest.
Symbolic regression recovers Kepler’s third law from real exoplanets, then quantifies what an interpretable formula costs against gradient boosting.
Hybrid BM25 + embedding retrieval over 3,100 papers, cited answers from a local LLM with automatic citation checks, and an offline evaluation harness.
Evidence-grounded marketing-brief risk triage with Amazon Bedrock structured outputs, versioned policy catalogs, strict allowlisting, and a synthetic offline evaluation harness.
02 / Capabilities
Production constraints shape the model from day one: auditability, failure modes, cost, and the people operating it.
Cleaning, normalization, validation, feature engineering, SQL, and reproducible Python pipelines.
Problem framing, baselines, supervised and unsupervised learning, calibration, evaluation, and diagnostics.
Document understanding, RAG, structured extraction, prompt versioning, validation, and evaluation.
Serverless components, storage, packaging, observability, infrastructure as code, and reproducible releases.
03 / Background
Training in physics, cloud ML certifications, and projects that report their limitations as clearly as their results.
Modeling, statistics and scientific computing: the base for treating every model as a hypothesis to test.
Comfortable presenting technical results to technical and non-technical audiences in both languages.
04 / Working principles
Define who acts, under which constraint, and what failure costs before choosing a model.
Probabilities, thresholds, rules, and human review are different layers—and should be testable as such.
A metric without context or an extraction without provenance is not ready for a consequential workflow.
05 / Contact
I am open to data science and machine-learning roles where reliable delivery matters as much as model quality.
manuelalejandropologonzalez@gmail.com