TL;DR
Built a modeling pipeline that predicts risk trajectories (how risk evolves over time) rather than a single static score. This is a better fit for interventions because timing matters.
Problem
Many risk models output one number (“high vs low risk”), but teams need to know:
- who is at risk,
- when risk is rising,
- which behaviors are driving the change.
Approach
- Data: longitudinal behavioral signals (multi-week windows) + optional survey/assessment anchors.
- Modeling: trajectory-aware risk modeling (sequence features + interpretable groups).
- Interpretability: summarize drivers per trajectory (what changed, not just how much).
- Evaluation: predictive metrics + calibration + subgroup/slice checks.
Results (what to add)
Replace this section with one chart and 2–3 bullet results:
- Baseline vs model (AUC/PR/MAE)
- Calibration (reliability curve)
- Example: trajectory segments that triggered an intervention
My role
- End-to-end pipeline: data cleaning → feature design → modeling → evaluation → write-up.
- Communicated trade-offs and deployment constraints for a product/ops audience.