IMRS: Production Multi-Label NLP for Incident Narratives

End-to-end NLP system replacing manual labeling of correctional incident free text with automated multi-label prediction across facilities.

Role: Technical lead / product owner (BIOS program) Partners: Recidiviz ; Council of State Governments

Highlights:

• 73 binary incident-type labels across 110,928 labeled incidents from 24-table relational export
• One-vs-rest MultiOutputClassifier with logistic regression; ongoing SVM experimentation (+~5 F1)
• k-fold cross-validation with Jaccard scoring on 27,700 held-out records
• Human-in-the-loop quality controls for continuous refinement

Outcomes: 75.6% micro-F1, 88.3% precision, 66.1% recall, 60.8% micro-Jaccard, 60.6% exact match; ~30x vs. random baseline

Python scikit-learn pandas TF-IDF Oracle/SQL NLP production ML