AI Lab · By Matt Zheng
Find the members who are both expensive and slipping away.
A care-team agent over a health-plan book of business. It flags at-risk members with transparent, reason-coded rules, drafts a coordinator note under a supervisor / QC layer, and keeps a human in the loop before any outreach. Everything below runs live in your browser.
The product insight: engagement alone does not predict cost. Average claims are roughly flat across engagement tiers, so a disengaged member is not automatically a cheap one. Flagging has to sit at the intersection of three signals: clinical risk, behavioral disengagement, and acute utilization.
Working dataSample data is loaded. Drop in your own CSV to flag a different population.
MembersSAMPLE
500 rows
columns: member_id, enrollment_date, plan_type, age_band, risk_score
ClaimsSAMPLE
1,109 rows
columns: claim_id, member_id, claim_date, claim_type, allowed_amount, diagnosis_category, orphan_member
Engagement eventsSAMPLE
12,594 rows
columns: event_id, member_id, event_date, event_type, session_duration_seconds, orphan_member
Runs entirely in your browser. Step 3 gives you risk and cost threshold sliders. Drag them and the flagged cohort recomputes live.
Built by Matt Zheng. The flagging engine is a deterministic TypeScript port of the Python agent, verified to reproduce the submitted result (32 flagged: 26 R1, 11 R2, 5 both) on the sample data. Care-team notes are real model outputs, drafted live on demand via MiniMax (M2 drafter / M3 supervisor). · Design notes