Your data never leaves the box. Pull the cable and prove it.
LastBeacon builds and deploys turnkey, fully-offline AI appliances for organizations that cannot — or will not — send their files to the cloud. One box. One tenant. A private model that answers from your own documents, cites the exact passage, and refuses to guess. The air-gap isn't a policy you take on faith. It's a test you can run yourself.
No cloudNo accountNo telemetryOpen-weight & license-clean
0%
on-prem · your hardware
0 B
outbound · default-deny
1
tenant per appliance
∞
offline · no signal needed
The proof we ship with every box
Watch nothing leave.
On the left, the appliance answers a question grounded in a sample confidential contract. On the right, an egress monitor shows every outbound packet attempt. Pull the network connector below — the WAN indicator goes dark, the monitor stays pinned at zero, and the answer keeps coming, cited to the document. Re-plug it; nothing changes, because nothing ever needed it. You don't have to believe us. You can disconnect us and check.
lastbeacon · local inference · demo
ask>
egress monitor · representative demo··:··:··
OUTBOUND0 B
PACKETS0
FIREWALLdefault-deny
WAN LINKconnected
STATUS · SEALED — 0 bytes left this machine
WAN
drag the plug →
Representative demo of the proof shipped with every appliance — not a live capture of your box. On a real unit you run it on your own hardware, with your own traffic monitor.
What we build
Two things, done properly.
01 · Offline model
Offline AI on hardware you own
We configure an open-weight model that runs entirely on-device or on-prem — no cloud, no account, no internet path out. Unplug it from the network and it answers exactly the same. The model is downloaded once at provisioning and never again.
02 · Grounded on your docs
RAG trained on your own documents
We ground the assistant in your material — contracts, case files, manuals, records, your knowledge base — so it answers from your files, not the open web. You point us at a folder; we ingest, tune retrieval, and wire the citation and abstain gates.
Running a model offline is the easy part — that's commodity. The value is that the answers come from your files, the air-gap is provable, and we deliver and support the whole thing. Turnkey, single-tenant, yours.
The interface
A chat your team already knows how to use.
It looks like the assistant everyone's used to — ask a question, get an answer. The difference is underneath: every reply is drawn only from your own files and carries a citation you can click straight to the source passage. The whole indexed corpus and when each document was processed are one click away, and the air-gap status sits in the corner. This is the actual on-box app, running fully offline.
Representative capture of the on-box assistant answering from a sample document set — local model, cited to the source, no network.
Autonomy, not rented access
No meter. No "buy more tokens." No paywall mid-sentence.
A cloud assistant you don't own can ration your usage, change its price, or retire the model you've come to depend on — sometimes in the middle of the job. You've seen it: the answer stops and a banner asks for your card. An appliance you own has no meter to run dry — nothing cuts it off mid-sentence. Same prompt, two business models — watch the difference.
NimbusAI · metered cloud · illustrativerented
50,000 tokens left
you> Summarise these 80 contracts and flag any unusual indemnity clauses.
! You're out of tokens.
0 of 50,000 monthly tokens remaining. Continue at $0.02 / 1K tokens — billed to the card on file.
Add credits →
lastbeacon · local inference · illustrativeowned
no token meter
you> Summarise these 80 contracts and flag any unusual indemnity clauses.
no usage bill · no per-token fee · keeps working offline
Illustrative comparison. "NimbusAI" is a fictional stand-in for the usage-metered cloud model — not a real product, and not a claim about any one provider. The point is structural: rented access can be metered, repriced, or retired mid-task; hardware you own can't be. The electricity and the one-time hardware are yours — there's simply no per-token bill, and no vendor who can cut you off.
See it in your world
What it actually does, in the rooms it's built for.
Same box, same promise — grounded in your own files, cited to the page, working with the cable pulled. Pick a world and watch it answer.
The moat — evidence, not assertion
Every answer points to the passage it came from.
There is no synthesized confidence and no anonymous summary. Each statement carries a citation to the exact supporting text in your own files — the document, the page, the verbatim quote. Click it and jump straight to the highlighted source. Your reviewers don't take the model's word for anything; they read the original, in context, in seconds.
Answer
Either party may terminate for material breach, but only after a 30-day cure period from written notice. Master_Agreement_v4.pdf · p.12
Source — your document
§ 11.2 Either party may terminate this Agreement for material breach upon thirty (30) days' written notice, provided the breaching party has not cured such breach within that period. Termination shall not relieve either party of obligations accrued prior to the effective date.
click the citation to jump to the supporting passage
Our honesty posture
When your documents don't support an answer, it says so.
We will not tell you the AI is accurate or that it won't hallucinate — no honest vendor can. What we will tell you is exactly how it behaves when it is unsure: two independent gates check whether the retrieved passages actually support a response. If they don't, the box returns "no source for this in your documents" instead of inventing one. You see the evidence and its strength, never a fabricated truth score. A human stays in the loop. Abstention is the feature, not the failure.
Compliance & autonomy architecture
Built to survive an audit. Owned, not rented.
One customer, one appliance, one private data store — no shared multi-tenant cloud, no neighbors, no commingled vectors. The models are open-weight and license-clean, so nothing phones home, nothing locks you in, and there is no remote disable. When the engagement ends, the box, the weights, and the data are simply yours.
Encrypted at rest
Single-box, single-tenant storage encrypted on disk. Your corpus, your vectors, your audit log — one private store, no neighbors.
License-clean, open weights
Apache-2.0 / MIT · Qwen3 · Mistral gte / bge embeddings · DeBERTa entailment on the support gate
No cloud dependency
No vendor account, no phone-home, no remote kill-switch. Autonomy scoped to what's falsifiable — not an uptime promise we can't keep.
Append-only audit log
Every query and every source returned is recorded so you can answer "who asked what, and what did it cite?" months later. (This proves the appliance is auditable — not that the model's outputs are correct.)
audit.log · append-only · hash-chained
Who it's for
For the rooms where data cannot leave.
If your answer to "can this go to the cloud?" is no, this was built for you. The box works the same on a mountain as in a SCIF.
Law firms
Privilege and client confidentiality that can't touch a third-party server.
scope a build →
Healthcare
Patient data bound by rules that a cloud assistant can't satisfy.
scope a build →
Finance & audit
Where the wrong outbound packet is a reportable event.
scope a build →
Government
Data-sovereignty mandates and residency requirements.
scope a build →
Field & defense-adjacent
Air-gapped by requirement, or working where there is no signal at all.
scope a build →
Individuals
Anyone who simply wants a capable AI they fully own.
scope a build →
Sized to the work
Every tier runs air-gapped. The only difference is headroom.
Entry
Proof of concept
Evaluate the grounding and the air-gap on your own files.
Compact Mac mini
Local model + RAG
Source attribution (L0)
Runs fully offline
Pro · recommended
The sweet spot
The configuration most teams keep — quiet, low-power, under a desk.
We recommend the tier; you own the hardware either way.
Turnkey delivery
From your folder to a working box.
There is no integration project for you to staff and no cloud account to govern. Request a build and we'll start with a confidential conversation, not a contract.
Scope
You tell us what must stay private; we scope the docs and the tier.
Harden
Provision the appliance and lock the default-deny firewall.
Ingest
Parse and chunk your material into a private index.
Tune
Tune retrieval against real questions from your team.
Prove
Run the egress proof with you before handover.
Support
We stand behind it and keep it working.
Based in Toronto — delivered on-site or shipped, anywhere in Canada and the United States.
Start a confidential conversation
Tell us what you can't afford to leak.
Send us your name, your organization, and what you need to keep private. We'll come back with the right tier, an honest read on what your documents can and can't support, and an egress proof you can run before you trust us with a single file. Your data never leaves the box. Let us prove it on yours.
This form posts to a Google Form we control to capture your name, email, organization, and message — nothing more. We don't ask you to upload documents to scope a build, and we don't drop this into a third-party CRM that resells data. The conversation starts here; your data never leaves your hands.
We'll reply to with the right tier, an honest read on what your documents can support, and an egress proof you can run yourself. No obligation. You can also reach us directly at info@lastbeacon.ca.
