A working portfolio of inventions across IoT, edge AI, and physical computing. Patents are written in legalese — the real stories are below.
US20020120728A1 — The First IoT System (2000)
Filed December 22, 2000. Inventors: Jason Braatz, Michael Hendricks, Eman Safadi. Method and apparatus for network-enablement of devices using device intelligence and network architecture.
Before the term "Internet of Things" existed — before the cloud existed — this patent described what the world now calls IoT. Any electrical device could be made network-addressable through a common protocol, controlled from a remote console, and rolled up into a server that aggregated state across an entire fleet. Not just read it. Control it. Listen to your stereo at home from your office. Adjust your thermostat from the airport. Turn off the lights from a hotel.
The architecture introduced two ideas that have since become foundational: a device-independent controller speaking a common protocol, and a device-dependent controller translating to whatever quirky native interface the actual hardware wanted. Above them sat an abstraction layer for both direct control and aggregated telemetry. Everything that lives in modern IoT — from Nest to industrial SCADA — descends from this shape.
Then Google noticed. Larry Page made my cohorts and me an offer we could not refuse: come join Google, build this from the inside, and let the application go abandoned so Google could be first into this space. We took the offer. The patent went abandoned in 2002 — on purpose. It remains the first formal definition of what is now called IoT, and it is cited across the research and patent literature that followed.
US12324122B2 — Portable Wood-Cooled Edge AI for Off-Grid Medicine (2024)
Granted 2024. Inventor: Jason Braatz. Portable autonomously powered computer with thermally separated heat energy sources in vorticity-inducing cavity.
The use case was specific. During the Middle East refugee crisis, Doctors Without Borders physicians were running 400:1 patient ratios in camps where insurgents actively hunted medical facilities. A modern medical SLM with full PubMed and JAMA RAG would help triage nurses pre-screen patients and route the most urgent ones to the doctor first. The catch: the hardware running it could not be detected. No radar signature, no lidar return, no drone-photograph thermal bloom, no fan noise, no EMI fingerprint. And it had to be portable. And it had to run a 12B-parameter model with a Milvus vector database underneath. Updates arrived by sneakernet — Ollama and Open WebUI on a USB drive, smuggled in.
The solution starts where most engineers stop: with wood. Most dismiss wood as a thermal insulator. It is — until you remember it is an anisotropic insulator. The gymnosperm channels that once moved water and sap through a living tree are oriented capillary structures. Separate the hot components carefully, route their heat through the wood grain rather than across it, and the same channels that carried sap now carry warm air. Cavity geometry induces vorticity, and the cavity becomes a passive heat exchanger. No fans. No pumps. No noise. No detectable thermal plume. To the radars and the drones, it looks like a piece of wood.
The result: a portable, fanless, undetectable computer with enough headroom to run the biggest local Llama you can quantize, fed by a RAG index over PubMed, JAMA, and the rest of the open medical literature. Triage nurses ran it ahead of the doctors. Doctors saw the right patients first. Outcomes improved. The chassis is wood, the silicon is GPU-class, and the story is what most papers about edge AI leave out: "production-ready" sometimes means "will not get the user kidnapped."
More to Come
Three hundred forty-four additional patents trace back to ML video protocol work at Google — many still embedded in YouTube infrastructure. Follow-on filings from Braatz Research are in process. This page is intentionally short and intentionally editable; as new inventions mature into filings, they will land here.