Projects
Applied Work | Open Projects | Structured Methods
Projects
Applied Work | Open Projects | Structured Methods
Applied Research Portfolio
Complete case studies demonstrating end-to-end research methodologies. Each project shows the full process: from problem definition through validation, with transparent documentation of decisions and tradeoffs.
What you’ll find here: Real research with real challenges, not curated success stories.
Featured Projects
Face Value
Weak Supervision Emotion Classification
Testing whether academic benchmark datasets (FER2013, RAF-DB) actually predict real-world performance.
The problem: Emotion recognition models trained on benchmark datasets often fail in real applications. Is this a modeling issue or a data issue?
The approach:
- Built emotion classifier using weak supervision with stock photos
- Compared three data sources (Pexels, Pixabay, Unsplash)
- Validated against movie timelines for ecological validity
- Face Value outperformed benchmarks as evidenced by less collapse and time series analysis
Key finding: Benchmark accuracy doesn’t predict real-world utility when domain mismatch exists.
Methods highlighted: - Weak supervision for dataset creation - Multi-source data quality comparison - Ecological validation strategies - PyTorch model fine-tuning
Pixel Process
Open-Source Platform Design & Content Engineering
Building an interactive data science learning platform and the tradeoffs involved in choosing the right delivery tool for each type of content.
The problem: Most data science education optimizes for completion rather than understanding. How do you build interactivity that actually helps people learn?
The approach:
- Designed three-tier interactivity system matching tool to content complexity
- Pyodide for zero-friction concept demos, JupyterLite for multi-step analysis, Binder for full ML workflows
- Built “Paths, Not Courses” navigation for non-linear learning
- Implemented intentional bugs as debugging practice
Key finding: No single interactive tool fits all content. Matching execution environment to learning objective matters more than maximizing interactivity everywhere.
Methods highlighted:
- Platform architecture decisions and tradeoffs
- Content strategy for technical audiences
- Quarto site development with custom CSS/JS
- Pyodide, JupyterLite, and Binder integration
You’re already on the site!
The entire site — Quarto config, custom styling, Pyodide cells, Binder integration, and all content — is open source. Fork it, learn from it, or build on it.
Looking for Custom Solutions?
These projects showcase open research and methodology. For custom research solutions, experimental design, and ML engineering consulting, visit Dexterous Data.