Chad Hurley, RN
I'm a med-surg float nurse at a Level 1 trauma center. Off the floor, I build AI that remembers. Its memory stays with the person instead of the model.
The handoff
In nursing, the most dangerous minute of a shift is the handoff. Everything the last nurse knew about your patient has to survive a five-minute conversation at the station. When it doesn't, people get hurt.
Most AI lives in a permanent handoff. Every new session starts from zero, and every model upgrade wipes the chart. I build memory that survives the handoff, so the engine underneath can change and the assistant still knows who it's talking to.
What the bedside taught me about building it:
- Date it, and never erase it. In a chart you don't delete a wrong entry. You correct it and leave the trail. My memory systems keep facts the same way: superseded, never deleted.
- Read back every critical number. A routine test caught one of my inference engines copying a code wrong once prompts ran past about 20,000 tokens. Short prompts were perfect. You only find that if you check.
- Know whose fact it is. A medication belongs to one patient. A memory system for a household has to know whose fact is whose, and refuse to answer when it isn't sure.
- Build for 3 a.m. Guards, runbooks, and a note for the next shift.
Nursing
- 2022 – nowMedical-surgical float nurse at a Level 1 trauma center on the Gulf Coast.
- 2021 – 2022Thomas Hospital, Infirmary Health.
- 2019 – 2021Oncology and neuro nursing at Ascension Providence. One of the first frontline nurses vaccinated for COVID-19 at Providence Hospital, covered by WKRG.
- AwardsThe DAISY Award for Extraordinary Nurses, four times, and a Nursing Excellence Award. DAISY nominations come from patients and their families.
AI that remembers
I design persistent-memory, platform-agnostic AI systems and run them on hardware I built: NVIDIA DGX Spark and Apple Mac Studio machines working as one fabric, with local models end to end.
The rules I build by:
- OwnerMemory belongs to the person, not the platform.
- RecordSupersede, never delete.
- NumbersRead back every critical number.
- PrivacyFail closed on anything private.
Projects
pd-bridge
Prefill a 284B model on NVIDIA and decode it on Apple Silicon, over plain 10GbE. Two inference engines that share no cache format, framework or vendor serve one request together. A 700,630-token prompt runs end to end in 11 minutes 26 seconds.
zigmem
A general-purpose memory system for local LLM agents. One Zig binary, one SQLite file, no cloud. Facts are superseded or retired with an audit trail, never hard-deleted, and household scoping keeps medical facts with the person they belong to.
Micro Agent
A small caretaker that runs before login and keeps a self-hosted AI memory stack alive. It guards, heals by runbook, refuses what it must never do, and leaves a note for the next session.
LaserMemory
Point it at a page, a PDF or text. A keeper decides what is worth remembering and shows its work.
TensorFold
Credited contributions to TensorFold, a fast, exact LLM engine for Apple Silicon.
Builder
I built and sold five companies before I went to nursing school, in environmental services, scrap metal recycling, car washes and landscaping, among them New Horizon Environmental, Heritage Landscape and Design, Minutemen Carwash and Grafton Carwash.
I hold a patent, and I designed and ran blockchain mining hardware before it was popular. I've been taking computers apart since the Commodore 64.
Family
Father of four daughters. Everything I build is for them.