ML-Powered Information Extraction
Product Design, UX Research, Prototyping
HURIDOCS
ML Engineers, CTO, PM
Figma, Miro
Applying machine learning to accelerate document review and entity extraction for human rights defenders.
Machine Learning, NLP, Information Extraction
Explored and prototyped ML-assisted extraction for Uwazi: surfacing key entities, reducing manual work, and validating accuracy with practitioners.
The Challenge
Human rights defenders deal with mountains of documents. Manually extracting names, places, and dates was slow and error-prone. The challenge was to explore whether machine learning could support their review process without overwhelming them.
Process & Research
Worked closely with our ML engineers to understand model outputs and failure modes. In Figma I designed and prototyped different ways of surfacing entities and tested how defenders reacted to suggestions and confidence scores.
Testing & Iteration
Through quick cycles with practitioners, I learned that trust hinged on transparency. Users wanted to know why a suggestion was made, and how sure the system was. I iterated on explanations, highlights, and feedback loops for training the model.
The Solution & Impact
Designed a workflow that allowed defenders to quickly accept, reject, or correct ML suggestions. This reduced manual effort while keeping humans in control. Accuracy improved over time, and confidence in the tool grew.
Learnings & Next Steps
Learned that ML is most useful when it fits into real human workflows. Next steps would include expanding extraction to more entity types and integrating feedback directly into model retraining.
