Behind Polvoria is a researcher. Alongside design and build work, I do academic research in machine learning, optimization, and computer vision — the same discipline, applied to measurable problems.
Intro film
A short introduction to the work.
I work at the intersection of machine learning, metaheuristic optimization, and applied data science — turning research methods into systems that hold up under measurement.
The throughline is discipline: a clear objective, a documented method, and a number to judge it by.
Open to
→Research collaborations
→Freelance ML & data science
→International roles — EU / Germany, PhD
Selected work
Measured results, not claims
Every project is grounded in a metric. Optimization is applied where it adds measurable value — never as a buzzword.
Intrusion detection · CICIDS2017
Network intrusion detection
LightGBM and Random Forest with GA feature selection and SHAP explainability.
98.14% accuracy · 98.6% macro F1
Feature selection · CIC-IDS2018
Genetic feature selection
A genetic algorithm via DEAP over 360K rows — complexity cut while accuracy held.
78 → 45 features · 42% reduction
Hyperparameters · CIC-IDS2018
PSO hyperparameter tuning
Particle swarm optimization tunes a LightGBM classifier toward a higher macro F1.
+0.26% macro F1
Segmentation · Remote sensing
Satellite image segmentation
A U-Net with GA multi-threshold optimization across 6 land-cover classes on 4-channel imagery.
6 classes · IoU tracked
In progress
Object detection · Automotive
Ford logo detection
A convolutional baseline moving to YOLO for real-time logo detection.
CNN → YOLO
Research & approach
One framework: optimize, then explain
Metaheuristics search the spaces gradient methods can't; machine learning does the prediction; explainability keeps every result accountable.
01 · GA
Genetic algorithms
Evolve feature subsets and thresholds — keeping what measurably helps and discarding the rest.
02 · PSO
Particle swarm
Tune model hyperparameters across a continuous space toward a single, stated objective.
03 · ML
Machine learning
Gradient-boosted trees and deep networks (LightGBM, U-Net) do the prediction.
04 · XAI
Explainable AI
SHAP attributes each prediction to its inputs, so a result can be read, audited, and trusted.
Accuracy alone is not enough. In security and remote sensing, a decision has to be defensible — SHAP turns a model from a black box into something a reviewer can interrogate, which is what makes the result usable in practice.