← all projects

RIPPLE: AI-Powered Vertical Supply Chain Management for Medicine

mapping the hidden single points of failure in the medicine supply chain

project: 04 of 05
event: DNHacks 2026
credit: Built over one weekend at DNHacks 2026 with four teammates.
links: Showcase → Source →

01the problem

most medicines depend on ingredients from a handful of overseas factories
six antibiotics that look diversified trace back to one shared precursor
if that factory stops, all six run short at once

six antibiotics, one precursor, one plant

02the graph and the risk model

public FDA registers pulled into one graph of 1,732 nodes and 3,425 edges
failure spreads through the graph by breadth first search with a redundancy rollup
an L2 regularized logistic model scores which plant fails next, from nine leakage free features
raw scores recalibrated so a displayed percentage matches observed frequency

The console tracing a drug through its suppliers and showing where they sit in the world. — Fig. 1. The console tracing a drug through its suppliers and showing where they sit in the world.

The model maps a weighted sum of features to a probability. — Eq. 1. The model maps a weighted sum of features to a probability.

The Wilson interval, used to check the displayed percentage against observed frequency. — Eq. 2. The Wilson interval, used to check the displayed percentage against observed frequency.

03rerouting and the sensor

switch off any factory and watch the disruption propagate
replacement suppliers ranked on capability, geography and their own risk
each checked against Trade Agreements Act and Defense Department restrictions
an ESP32 sensor streams cold chain telemetry and computes mean kinetic temperature

a plant goes down and supply routes around it

Ranked replacement suppliers, each scored and checked for procurement eligibility. — Fig. 2. Ranked replacement suppliers, each scored and checked for procurement eligibility.

the cold chain sensor node and its probe

Mean kinetic temperature, weighted by the time each reading represents. — Eq. 3. Mean kinetic temperature, weighted by the time each reading represents.

04results

model reached 0.814 AUC against a 0.70 baseline, beating it on all three time cutoffs
gradient boosting was tested and lost at 0.641, so the transparent model is also the accurate one
supplier name matching scored 100 percent on a labeled benchmark
covers 1,706 of 1,732 nodes, and the 26 it cannot resolve are left blank deliberately