Knowledge & retrieval
9 skills- Web fetch, search and page summarisation
- Semantic search across every past conversation
- Reading PDFs, spreadsheets, documents
- Video transcript extraction
- Reference lookup and translation
Loading
A private AI that runs entirely on hardware we own.
Sahr is a personal AI assistant our founder built for himself. It is deliberately not a product: single user, no cloud dependency at inference time, no subscription, and no plan to sell it. It runs on hardware sitting in a room in Islamabad and is reachable from a phone at any hour.
We publish it because it is the most honest evidence we can offer of what we can build. Most studios describing themselves as AI companies are calling a hosted API. Sahr required fine-tuning a model, building an evaluation harness strict enough to block our own releases, running inference on a constrained edge device, and keeping a memory system coherent across months.
Everything below is architecture-level on purpose. Network topology, service names and anything held in its memory stay private — it is a live system holding one person's data, and publishing an attack surface would be a strange way to demonstrate competence at privacy.
Status: In development · v1 · figures are point-in-time
Each skill is a separate contract the model can invoke, with its own failure mode and its own rows in the training corpus teaching when not to reach for it.
WithheldTen further skills are deliberately unlisted. Anything touching credentials, finances, physical location or command execution stays private — naming those on a public page tells an attacker exactly what is worth going after, which rather undermines the point of building it this way.
Jetson Orin Nano, 8 GB shared memory
Always on. Hosts the API, the memory store, and every skill. Answers from the phone at any hour.
Under 10 W at idle — it can run continuously without thinking about the electricity bill.
Single consumer workstation GPU
Model training, image and video generation, and heavy inference. Wakes on demand and suspends after a period of idleness.
Idle draw fell from roughly 70–90 W to about 30 W once the memory-release chain fired before suspend.
Apple Silicon peripheral
Streaming speech synthesis.
Explicitly temporary — the plan is to retire it once a larger edge device can handle voice locally.
Sahr has a body being built for it: 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 ruleThe 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.
Seven states, each with its own expression — procedural eyes that blink, drift and breathe rather than a looping animation.
Five layers, strictly downward-depending. Anything lower talks upward only by publishing an event — which is why the battery manager has never heard of the face renderer, and why new hardware slots in without touching the core.
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 now. 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 isRunning on real hardware. The board boots and the face is alive on its own display — eyes that blink, drift and breathe across nine emotional states. This is a working device on a bench, not a render of a product that does not exist. The phone link, end-to-end conversation and power hardening are still ahead, and the framework is layered so those slot in without rearchitecting what already works.
The chat model is a fine-tuned 7B, trained in-house rather than prompted. Every candidate is scored against frozen tool-use and behaviour suites, and only gets promoted if it beats the incumbent. One recent build was held back despite improving on conversational quality, because it did not beat the incumbent on tool use.
Why it mattersThis is the discipline missing from most AI work. When we tell you a feature hits an accuracy threshold on your data, it is because we built the harness that proves it — the same one that has blocked our own releases.
Most skills carry explicit restraint instructions — call this only when actually asked — and the training corpus contains rows teaching the model to decline. There is also a recovery path for when it routes to the wrong tool anyway.
Why it mattersAn agent is only as good as its worst tool call. Restraint is harder to teach than capability, and it is the difference between a demo and something you let touch production.
The always-on brain runs inside 8 GB of shared memory on an edge device, with heavier work handed to a single workstation that sleeps when idle and wakes on demand.
Why it mattersIf a model behaves inside 8 GB at the edge, it will behave inside your on-premise server. Cost and privacy stop being in tension.
Tiered memory with vector recall over a local embedded database — working, short and long-term tiers, promoted and decayed on their own schedule rather than dumped into one growing context window. A nightly consolidation pass reorganises what was learned that day.
Why it mattersThe reason most assistants feel amnesiac is that retrieval is an afterthought. Tiering is what keeps recall useful past the first fortnight.
A verified defence against confabulated tool use: the model can no longer describe having run something it never invoked.
Why it mattersAn assistant that says it sent the email when it did not is worse than one that cannot send email at all. This class of failure is the one that destroys trust fastest.
Speech recognition and streaming synthesis run on owned machines, including capture that transcribes a recording and files it into memory automatically.
Why it mattersVoice interfaces usually mean shipping audio of your meetings to a third party. It does not have to work that way.
Run against a scratch clone before anything is promoted, with pre-existing failures tracked separately so a real regression is never lost in the noise.
Every architectural decision numbered and recorded with its reasoning. Seventy-plus entries. When a choice is revisited two months later, the argument is still there.
A gotchas file capturing each way the system has broken and what fixed it — including the night a background process suspended a training run mid-step and brought it back as a zombie.
Push-to-deploy with the ability to revert cleanly, because a personal assistant that is down is a personal assistant nobody trusts.
The current system is best described as a responsive assistant with a persistent identity. It answers, remembers, and holds a consistent voice across months.
What it is being built toward is different: continuous presence. Something that carries on thinking between conversations rather than idling until spoken to.
Voice is currently output-only. Inbound speech recognition is the single largest step toward presence per hour of work — deliberately push-to-listen rather than an always-on microphone.
No physical-world sensing today. The target is knowing someone has walked into the room without being told.
Right now the gap between nightly consolidation passes is dead time. The goal is arriving at a conversation having already chewed on something from the last one.
Mood currently drifts in reaction to events. The next step is internal evolution, and a forward thread of its own — something it is working on rather than waiting on.
Non-goals
Evaluation is the product. Without a harness that can block a release, “the new model feels better” is your only signal — and it is wrong often enough to matter.
Constrained hardware forces honest engineering. There is nowhere to hide a wasteful design inside 8 GB.
Memory is a retrieval problem long before it is a model problem.
Teaching a model when not to call a tool is harder, and more valuable, than adding another tool.
A system that confabulates its own actions is worse than one with fewer capabilities. Grounding what it claims to have done matters more than what it can do.
The operational work — deploys, rollback, a written record of every failure — is most of the distance between a demo and a system.
Sahr itself stays personal. One user, one person's data, not for sale — that constraint is what made the engineering honest, and removing it would change what the thing is.
What we intend to offer is the capability, not the assistant. The same shape — a private model on hardware you own, optionally with a companion device — built for you, holding your data, 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.
Fine-tuned on your documents and your language, scored against a held-out set built from your own data rather than a public benchmark.
Some contexts do not want a phone in your hand — a workshop floor, a ward round, a driver's seat. The pendant work exists because we wanted to find out what that takes.
Be clear about thisTo be explicit: none of this is a product you can buy today. The private-AI engagement is real and available now as client work. The device is R&D, and we will not pretend otherwise or take a deposit against a date we cannot commit to.