Renato J. Quiliche Altamirano

Time-series forecasting and decision analysis for a wind-farm investment

Replicated the real-options valuation of Bastian-Pinto et al. (2021), where a Brazilian wind farm can sell its early energy or mine Bitcoin, and verified the port path-by-path against the authors' code. Replaced the paper's GBM with a bagged power-law + halving-cycle forecast (STL, block bootstrap, mean reversion) and scored both against the realized 2019–2025 prices.

Three models gave three different investment recommendations; the realized path shows why the starting-level assumption mattered more than the stochastic process.

• 0.25CRPS (vs 0.38)
• 96%90% coverage
• 50KMC paths

• R
• Monte Carlo
• Real options
• Bagging
• Shiny

Read the write-up → Original paper ↗

Scaling an M.Sc. dissertation model into a production-shaped Spark pipeline

Re-engineered my dissertation modelling stack as a distributed pipeline on Databricks, moving from single-node Python to Spark MLlib over 101K household records and 112 features. The pipeline chains VectorAssembler → RobustScaler → GBTClassifier, with cost-sensitive class weighting for a 1.4% positive rate and hyperparameter search via TrainValidationSplit.

Evaluation is imbalance-aware and computed with distributed Spark SQL aggregations (MCC, G-Mean, balanced accuracy, GIBA_α). The fitted PipelineModel is persisted to Unity Catalog Volumes for reproducible reload and batch scoring.

• 0.81AUC
• 0.67Sensitivity
• 0.83Specificity

• PySpark
• Spark MLlib
• Databricks
• Spark SQL
• Unity Catalog

M.Sc. dissertation codebase — PUC-Rio

End-to-end supervised learning pipelines and spatial data processing to model and predict household disaster risk in Peru, combining national household surveys with geoprocessed hazard data. Covers feature construction, hyperparameter optimization and evaluation under severe class imbalance.

Published in the International Journal of Disaster Risk Reduction (2023, 2025).

• Python
• scikit-learn
• Geoprocessing
• Imbalanced learning

Code ↗ Paper ↗

Automated evaluation for an ontology-grounded customer-service chatbot

Open-source experimental framework for evaluating LLM-based customer service agents against formal ontologies. Automated experimental pipelines in Python integrate large and small language models with handcrafted and automatically induced ontologies, plus retrieval augmented generation, running on GPU infrastructure.

• LLMs
• Formal ontologies
• RAG
• Python
• OntoUML

Code ↗

Spatial COVID-19 risk mapping for economic reactivation

Geoprocessing of COVID-19 deaths and affected populations combined with spatial econometrics to identify local contagion patterns, supporting a framework for phased economic reactivation. Published in MDPI Sustainability.

• Spatial econometrics
• Geoprocessing
• Python
• R

Paper ↗

Interactive analytics for research group performance

Custom interactive dashboards tracking the main indicators of researcher and student performance and scientific production, plus automated information collection and interpretation built on Google Workspace.

• Power BI
• Streamlit
• Python
• Google Workspace

UNDP & Government of Brazil

Designed and implemented policy surveys targeting municipal, state and federal public officials to assess the quality of risk and disaster management policies, in support of the Plano Nacional de Proteção e Defesa Civil. Combined descriptive analysis of survey data with natural language processing of interview transcripts.

• Survey design
• NLP
• Python
• Policy evaluation

Teaching assistance — PUC-Rio

Practical learning tasks for training and evaluating machine learning models, used in the graduate data science course at PUC-Rio.

• Python
• Machine learning
• Teaching

Code ↗