Pixel Process
Pixel Process
Think Clearly | Build Carefully | Apply Rigorously
Applied projects, research tools, and interactive methods for data scientists who question assumptions.
Featured Projects
Ensemble Methods · Computer Vision
ConserVision: Wildlife Camera Trap Classification
Rank 11 / Top 2% — DrivenData ConserVision Competition
Image classification competition on 8 categories (including blanks) with log-loss metric. Solution used 16-model ensemble pipeline trained entirely on a single RTX 2060. No exotic hardware, no proprietary data — just a well-designed pipeline and disciplined ensemble strategy.
Core Process: - Detection threshold with variation for more decorrelated predictions - Multiple fine tuned models with varying architecture and backbones - Ensemble testing with multiple models for optimized performance
Weak Supervision · Computer Vision
Expressions Ensemble
Testing whether academic emotion recognition datasets actually predict real-world performance. A complete weak supervision pipeline comparing Pexels and Pixabay data against FER2013 and RAF-DB.
Key finding: Aggregate and timeplot evaluation on 50+ full movies demonstrate how my weak supervision approach provides superior domain alignment than benchmark datasets.
Read the full case study →
Full Repo →
Interactive Results Dashboard →
Open-Source · Platform Design
Pixel Process
Experiences from developing an open-source learning platform about content delivery, interactivity, and finding the right tool for each problem.
Key finding: The right tool depends on where the learner is and what you’re trying to teach. There is no perfect solution.
Workflow · Foundations
Explore More
Workflow — Practical patterns for data science work that scales
Foundations — Interactive Python environments from zero-setup to full notebooks