Expiry

Project, 2026

Expiry radar of treatments medicine reversed, each showing what was standard and what guidelines say now
Timeline
2 days, Jun 2026
Tools
Next.js, TypeScript, Microsoft Foundry IQ, Vitest, Playwright, Vercel

Overview

Expiry flags treatments whose evidence a newer guideline has already reversed, so a clinician does not act on yesterday's standard of care. It covers ten real reversals across six specialties and scans a synthetic cohort of 60 patients against all of them. I built it for the Microsoft Agents League hackathon, Reasoning Agents track.

How it works

The rule is LLM at the edges, deterministic in the core. A pure, tested rules engine decides whether a reversal applies to a patient, so the same patient always gets the same flags and a reviewer can trace why. Flagging makes no LLM or network calls, and a cost-guard test fails if it ever does. The model only works when a clinician opens a single flag and asks for the reasoning.

Guideline knowledge baseMicrosoft Foundry IQRadarold vs new advice, the trial that flipped itDeterministic engineassessPatient / assessCohort: no LLMFlags + evidence timelineGrounded explanation, on demandone Foundry IQ retrieval, with citations
The pipeline. If Foundry is unavailable, the explanation falls back to the deterministic rationale.

Results

Radar, the cohort worklist, then Patient Lab, where you build a patient and watch the flags update.

In the demo cohort, 25 of the 60 synthetic patients are flagged and 35 are clear. The repo has 186 passing unit, contract and cost-guard tests plus Playwright smoke tests. Every patient is synthetic, and the repo holds no real patient data. Open the live app.