The Problem
One late night in Atlanta in the summer of 2025, the Trammell came out of a workout class to find his truck's window shattered and roughly $5,000 in technology missing. The thieves used a $10 center punch from a hardware store. He was less than 100 feet away the entire time and had no idea it was happening.
The incident exposed a gap in the market. According to Atlanta Police Department data, there were 5,711 Theft from Auto cases in the city between May 1, 2025 and May 1, 2026… existing security products aren't built for this scenario. Consumer glass break sensors from major brands all share the same limitations: they require a functioning WiFi network, a proprietary hub, a monthly monitoring subscription of $20–50, and a fixed wall-mounted installation. A parked car in a garage, on a street, or in a friend's driveway has none of those things.
The Requirements
Four requirements defined the build from day one:
1.Alert from any distance: the owner must be notified immediately, no matter how far away they are.
2.Fully portable — no hardwiring, no wall anchor; movable between vehicles.
3.Deterrence — the device should let intruders know they've been detected.
4.Low power — long-duration operation on a battery pack.
Early concepts were considered and rejected. Window-mounted vibration sensors failed the portability requirement. PIR motion sensing had recently been produced by others. The winning concept: a continuously listening device, like a baby monitor with a brain, that classifies sound in real time and pushes an alert the instant it hears glass break.
The Solution
G.L.A.S.S. is a standalone, battery-powered IoT device that performs real-time audio classification entirely on-device and delivers push alerts over a cellular LTE connection. No WiFi. No hub. No subscription.
Edge AI Pipeline
Rather than training an audio model from scratch, which would require huge labeled datasets, the system uses transfer learning:
Stage 1: A quantized TFLite version of Google's YAMNet model, pretrained on the AudioSet dataset of over 1.5 million audio clips, extracts a 521-dimensional feature vector from each ~1-second audio window.
Stage 2: A lightweight custom classifier, trained on labeled examples of target sounds, maps that feature vector to a class label and confidence score.
Both models combined weigh in at roughly 50 KB and run entirely on the device. No audio ever leaves the unit, a privacy guarantee that cloud-connected competitors can't make.
Cellular Alerting
Connectivity comes from an LTE-M/NB-IoT modem purposely built for low-power IoT applications. When a target sound is detected above a specified confidence threshold, the device publishes an alert over the cellular data connection to a cloud notification service, and the owner's phone receives a high-priority push notification within seconds. Because alerting rides on cellular infrastructure that already blankets the entire country, the device works identically in a parking deck, on the street, or in a rural driveway... anywhere with signal.
What's Next
Development is underway on production hardware, a companion mobile application, and expanded deterrence capabilities. The core system architecture is the subject of a pending patent application.
G.L.A.S.S. is developed by Montgomery Research Labs LLC. For inquiries, partnership, or updates, visit montgomeryresearchlabs.com/g-l-a-s-s . Patent Pending — the system architecture described herein is the subject of a provisional patent application filed with the USPTO.