Projects

Selected research and technical case studies in computational biology, machine learning, forecasting, computer vision, and signal processing.

Selected work

Research questions turned into reproducible evidence.

My work spans computational biology, applied machine learning, forecasting, computer vision, and signal processing. Each case study emphasizes a clear validation question, transparent metrics, and limitations that matter beyond a single model score.

6documented case studies
3technical focus areas
R + Pythoncross-language research workflows

Research & Computational Biology

Multi-omics integration and group-aware scientific modeling
Computational biology Active research

Joint-RPCA for Multi-omics Integration

Cross-language validation of R/Bioconductor and Python workflows for microbiome data.

60 shared samples retained across the current multi-omics workflow
  • Aligns filtering, configuration, and randomness across R and Python implementations.
  • Compares scores and loadings after accounting for latent-space indeterminacy.
R Python Bioconductor Gemelli Quarto
Case study R implementation mia development Workflow Gemelli
Predictive modeling Completed study

Protein–Drug Affinity Prediction

Validation study showing how evaluation design changes apparent model generalization.

≈0.512 concordance index when generalizing to an unseen protein
  • Contrasts observation-, protein-, and drug-level holdout strategies.
  • Demonstrates how a strong aggregate score can conceal weak deployment generalization.
Python scikit-learn grouped CV concordance index
Case study Grouped CV

Machine Learning

Generalization-focused predictive and computer-vision systems
Time-series forecasting Completed study

Solar Energy Forecasting

Statistical and machine learning models for short-term photovoltaic production forecasting.

≈0.1618 best non-differenced XGBoost MAE in the project target scale
  • Benchmarks statistical and machine learning models on identical forecast windows.
  • Preserves temporal ordering to prevent future-to-past leakage.
Python R XGBoost ARIMA/ETS time-series CV
Case study TimeSeriesSplit XGBoost
Sensor classification Completed study

Human Activity Recognition

Participant-aware classification with random forests, support vector machines, and LOSO validation.

LOSO participant-level validation for performance on unseen users
  • Compares random forests and support vector machines with consistent model selection.
  • Keeps every participant entirely within one validation fold.
Python scikit-learn random forest SVM grouped CV
Case study LeaveOneGroupOut
Computer vision Completed study

RGB and NIR Image Segmentation

Controlled comparison of RGB and multimodal inputs for semantic segmentation.

87.01% RGB test score in the controlled input-modality comparison
  • Compares RGB with combined RGB and near-infrared inputs under one protocol.
  • Shows that an additional modality does not guarantee improved generalization.
Python deep learning RGB/NIR semantic segmentation
Case study Segmentation models

Data Science & Signal Processing

Forecasting and physiological signal analysis
Signal processing Completed study

ECG Heart-rate Estimation

Signal-processing pipeline for robust heart-rate estimation from short ECG segments.

200 × 30s ECG segments analyzed at a sampling rate of 200 Hz
  • Combines peak detection with autocorrelation and spectral evidence.
  • Makes low-confidence segments and physiological plausibility checks explicit.
Python NumPy SciPy peak detection spectral analysis
Case study Peak detection Welch PSD

Validation matches the question

Temporal, participant-aware, and group-aware splits are chosen to reflect the intended use of a model.

Reproducibility is engineered

Preprocessing, seeds, configurations, software versions, and intermediate outputs are made explicit.

Limitations stay visible

Results are presented with their dataset scope, uncertainty, and external-validation requirements.

Public code links are shown only where a matching repository is available. Other links lead to authoritative method documentation used to contextualize the case study.