BrainLow power, always on
The always-on node: this node is Sahr
Jetson Orin Nano Super, 8 GB
- Identity and persona
- Long-term memory
- Mood and the dream cycle
- Skill dispatcher
- Safety gates
Loading
L01 · internal R&D · not for sale
A private, single-user AI that lives on hardware we own.
Idle
Waiting. Breathing, blinking, looking around.
Try her expressions
One Sahr, every surface
Transcript of the animation.
A sample turn, drawn as an illustration. Scores and probabilities in the film are not measurements.
The one-paragraph truth
Sahr is a single-user personal AI that our founder built for himself. She is not a product, not for sale and not multi-user, and she never leaves owned hardware. Her identity lives on one always-on node: her persona, long-term memory, mood, a nightly dream cycle and her safety nets. Every other machine is a service provider, not a brain. The design bias is local-first and privacy-first, with the cloud used only by choice, for specific bursty jobs.
We publish her because she is the most honest evidence we can offer of what we can build. Retrieval over dozens of skills, a decision API we trained ourselves, memory that consolidates overnight, and safety gates that would rather say "I can't pull that live" than invent an answer. All of it runs on machines we own.
Everything here is architecture-level on purpose. Network details, service names and anything held in her memory stay private. She is a live system holding one person's data, and publishing an attack surface would be a strange way to show competence at privacy.
by design, and not for sale
a few retrieved for each turn
composed for each turn
where her identity lives
a mixture-of-experts chat model
The fleet
Four owned machines with different jobs. One of them is Sahr; the other three serve her. No network details are published.
Jetson Orin Nano Super, 8 GB
RTX 5090 workstation
DGX Spark, GB10, 128 GB
Apple Silicon Mac
The rule. Move the service machines far away and Sahr is still Sahr, just thinking slower. Nothing assumes a GPU is always awake: the forge sleeps, nodes can be down, and she degrades instead of breaking.
What she can do
About 52 skills, but the model never sees them all. A dispatcher retrieves only the few that fit this turn, so the choice stays small and the answer stays fast.
Fourteen composable slices are put together for each turn: owner or guest, voice or text, and her current mood. Warm and concise, and at home switching between English and Roman Urdu mid-sentence.
Recall by embedding similarity, plus a nightly dream cycle that turns the day's raw turns into durable memories. A mood model and a growth journal sit alongside.
An owned decision API. Fine-tuned small encoders return calibrated yes/no probabilities for safety calls such as prompt-injection screening, sensitive-action gating and memory relevance, in one fast local pass.
Guards that fail closed, a reviewer gate on consequential actions, a groundedness check that tests whether the answer is supported by what the tools returned, and anti-fabrication scrubbers for when no tool ran.
A web console, iOS and Android companions, a WhatsApp bridge and a hold-to-talk voice loop: speech in, chat, speech out. She also has owned browser automation, behind a security guard.
The chat model
About 3B parameters active out of 35B, served from the owned capacity node with thinking off by default for latency. Embeddings are computed locally.
Some of the skills
Withheld. Some skills are deliberately not named. A public list of everything a private system can reach is a map for whoever wants to attack it.
Recently built
Hybrid lexical and embedding retrieval is live. An applicability rerank on top of it is built and being evaluated before it is switched on.
Scrubbers that replace invented news or clock times with an honest "I can't pull that live" when no tool fired.
An inbox digest and triage skill on a local index. Read-only, and owner-only.
Moving to a newer base model after it won a measured bake-off against the old one.
Telemetry, per-node health history and a synthetic-user simulation sandbox, to watch her real behaviour safely.
Discipline
New behaviour ships switched off behind a flag, is shadow-measured on a golden set, and is enabled only if it beats the baseline. A NO-GO result is respected, not argued with.
Capabilities are built in-house and run locally wherever that is feasible. Third-party clouds are opt-in, for specific jobs only.
She never fabricates tool output or data. An empty, honest "couldn't pull it up" beats a confident guess.
No multi-user layer, no tenants, no enterprise edition. Everything is for one person, on owned hardware.
Plainly
Ongoing work
Sahr has a body being built for her: a wearable pendant. Thin, low-power, and deliberately stupid. It renders an animated face, captures intent, and speaks one small protocol to a phone. Every ounce of intelligence stays off the device.
That constraint is the whole design. A pendant that tried to think would need a battery you would notice, and would leak your conversations to whoever made the chip. This one is a terminal: it shows you a face, listens when you press it, and forwards the question to a machine you own.
The rule. The brain is Sahr, not a third-party API. The firmware neither knows nor cares which model answered. It stays dumb, fast and gorgeous, and the intelligence lives on hardware you control.
What it is made of
States it shows
How the firmware is layered
Five layers, each depending only on the ones below. A lower layer talks upward only by publishing an event, so new hardware slots in without touching the core.
Built in stages
Prove the hardware and the render loop: a static face drawing at a stable 30 fps on real glass.
Procedural eyes, blinking, gaze drift and breathing, with nine emotional states morphing between one another, running on the display. Tuned on real glass rather than in a simulator, because this part is judged by eye and nothing else.
Bluetooth pairing, reconnection, and a small protocol so a phone can drive the pendant through its states while a waveform tracks the audio level.
Press, speak, and hear an answer: captured on the phone, sent to the brain you own, spoken back. The face reflects every phase of it.
Power and reliability hardening, which is where most wearables quietly fail.
Where it actually is. Ongoing work on a bench. The board boots and the face is alive on its own display, with eyes that blink, drift and breathe across nine emotional states. The phone link, end-to-end conversation and power hardening are still ahead.
Lessons
Evaluation is the product. Without a gate that can block a release, "the new version feels better" is your only signal, and it is wrong often enough to matter.
Show the model less. Retrieving the few skills that fit a turn beats handing it all of them.
Small, owned, calibrated models make better safety switches than a large model asked nicely.
Memory is a retrieval and consolidation problem long before it is a model problem.
An assistant that invents a result is worse than one that admits it could not get one.
Keep the identity in one small place. Everything else can sleep, fail or move.
The part you can buy
Sahr herself stays personal. One user, one person's data, not for sale. That constraint is what kept the engineering honest, and removing it would change what she is. What we offer is the capability, not the assistant. The same shape, a private model on hardware you own, built for you, holding your data and answering to you.
The deployment sits in your building or on your device. No inference call leaves your network, which is the entire point for anyone whose data legally cannot.
Models chosen, and where it pays off fine-tuned, for your documents and your language, then scored against a held-out set built from your own data rather than a public benchmark.
The same discipline that runs Sahr: behaviour ships off, is measured against a baseline, and is switched on only when it wins. Honest refusals over confident guesses.
To be clear. To be explicit: none of this is a product you can buy today. The private-AI engagement is real and available now as client work. Sahr and her pendant are R&D, and we will not pretend otherwise.
The same shape as this project, a private model on hardware you own, holding your data and answering to you, is available now as client work.