on-device
Local LLMs on device
Quantized models running on phones and edge hardware with embedded vector stores. Private, offline, instant.
Las Vegas · local-first · the workshop is open
AI you can
run yourself.
I'm Rich Olson. I run language models on phones, agents on persistent loops, and all of it on hardware I rack myself — and this site teaches you to do the same. Live exhibits you can touch, tools you can download, field notes from the edge of local and frontier AI. Nothing for sale.
US Patent 12,609,974 · granted 2026 · sole inventor
400+ weekly npm downloads
Live demos below · zero tokens leave the page
harness · on-device
// what I build
The workshop bench.
Six tools I reach for, end to end — from a quantized model running offline on a phone to a patented live-video pipeline.
agentic
Agentic harnesses
Multiple LLMs on persistent task loops with custom tooling — in production apps and in test rigs for features still in development.
retrieval
RAG & vector search
Custom embeddings over your data, retrieval that cites its sources, tuned to run where the data lives.
real-time
Streaming & live comms
Patented live-video technology, real-time rooms, RTSP pipelines and low-latency delivery that holds up.
native
iOS & Android
Web-view apps calling native via the Capacitor bridge — background downloads, full offline, Socket.IO across apps.
platform
Hosting · AtomicFalls
My own prompt-to-production platform — describe the site, and it's built, hosted, and live on infrastructure I run myself. Build it, host it, ship it.
// featured
Live video, on patented rails.
PATENT 12,609,974 · LIVE STREAMING
Project Vara Video
A private, browser-based conferencing platform — Zoom-scale in scope, friction-free in practice. Built on streaming techniques I hold a US patent on — granted 2026, sole inventor.
- Browser-based rooms with frictionless guest access
- Screen sharing & live file exchange
- Encrypted in transit and at rest
- Built on patented live-streaming methods
// open source
Pulled straight off the bench.
Tools leave the shop when they're ready — documented, tested, and worth your time. Three are out now. More are on the bench.
capacitor-socket-io
A native Socket.IO client plugin for Capacitor — robust, CORS-free real-time comms on iOS and Android.
JavaScript
2
codex-thread-tools
Tools and a Codex skill for keeping long agent threads healthy — health checks, handoffs, recovery.
Python
2
checkStream
Async check for whether an RTSP stream is live, using ffprobe to read codec and duration info.
JavaScript
2
···
more coming out of the shop →
// my favorite problem
On-device RAG, no round trips.
My favorite work: local LLMs running on phones and devices, answering from custom data embedded in vector databases. Private, offline, instant.
Try it — this harness runs entirely in your browser. No network, retrieving from a tiny store of facts about me.
100% client-side
0 tokens leave the page
ask-the-bench · v1● local
local store:
— run a query to retrieve —
answer:
Ask about the patent, on-device RAG, VARA, Capacitor apps, or hosting.
// who I am
Built to ship. Pointed entirely at AI.
I run my own DNS, mail, video, hosting, and inference — on hardware I rack myself — because intelligence you depend on should be something you control.
If it's on this site, it's running. Nothing here is a mockup.
REAL-TIME DISCIPLINE, APPLIED TO AI
- ✓Tight feedback loops
- ✓Observable behavior
- ✓Failure modes designed before launch
100%
self-hosted — DNS, mail, video & inference on my own iron
1
US patent, sole inventor — live video streaming
24/7
real-time systems and operational tooling
0
tokens leave this page in the live demos
// how I work
The bench is full. On purpose.
heads-down on my own AI projects
direct — no intermediaries
one exception: interesting problems
I'm heads-down on my own AI projects — a few you can see in the exhibits here, and a few I'm not talking about yet. I'm not taking client work, with one exception: a problem interesting enough to earn a spot on the bench.
If yours is that problem: the person you email is the person who architects it, builds it, and ships it — no account managers, no hand-offs, no juniors learning on your dime.
GOOD FIT
- ✓A defined problem where AI is the means, not the pitch
- ✓A proof of concept that needs to actually prove something
- ✓A custom AI-centric product built by someone senior
- ✓A decision-maker who wants a working session, not a slide deck
NOT A FIT
- ✕"Add AI to our thing" with no problem statement
- ✕Staff augmentation
- ✕Anything that needs a team of ten
Not right for me? I'll tell you quickly — and where I can, point you to someone who is.
// the one exception
Make the case.
The bench is full and I'm not looking for work — but a genuinely interesting problem can still earn the exception. Tell me why yours is the one. It comes straight to me, and I'll tell you quickly either way.
Las Vegas, NV
In Vegas? I do my best work across a table — coffee's on me.
github.com/zenzig
atomicfalls.com · hosting