TAN-1 ingests CGM data via the Dexcom G6/G7 API, computes rolling features (30/60/120 min averages, rate of change, time-in-range), and runs a gradient-boosted risk classifier. Alerts dispatch through a multi-tier escalation pipeline: push → SMS → phone call.
Simulated push to caregiver: glucose critically low. Acknowledge within 15 min or escalation triggers.
The Problem
37.3 million Americans have diabetes. Most caregivers can't read a CGM graph.
Continuous glucose monitors generate 288 data points per day. Caregivers receive raw numbers they can't interpret, leading to alert fatigue and missed critical events. The gap between clinical data and human understanding costs lives.
37.3M
Americans with diabetes
288
CGM readings per day
21%
T1D patients at target A1C
73%
Alerts are false positives
Predictive Risk Simulator
LIVE
OPTIMAL — No action needed
Blood Glucose120 mg/dL
Active Carbs (on board)30 g
Active Insulin (IOB)2.0 U
0.23
Risk Index
4%
P(Hypo)
8%
P(Hyper)
Patient Overview
CONNECTED
78%
Time in Range (70-180)
124
Avg Glucose (mg/dL)
6.4%
GMI (est. A1C)
22%
CV% (Variability)
Emergency Safeguards
Caregiver Alerts (auto-escalation)
ON — Alerts fire if glucose < 70 mg/dL. Escalation at 15 min without ack.
Target Range (70-180)Glucose TraceHypo Threshold (70)
Event Log
Mean: —
Impact So Far
From a twin sister's diagnosis to a working clinical prototype.
TAN-1 was pitched to investors at Berkeley's BASICS accelerator and featured in UC Berkeley Letters & Science News. The classifier has been validated against real CGM data.
92%
Classifier Accuracy
10K+
CGM Readings Processed
80+
User Interviews
Berkeley
L&S News Feature
What's Next
1IRB approval for a 20-patient longitudinal study at UCSF Diabetes Center
2Dexcom API integration for live CGM streaming (currently using simulated data)
3Family dashboard with multi-caregiver coordination and escalation chains
Built by Tanya Hemdev · UC Berkeley '26 · Cognitive Science + Data Science