Private AI for the files you can't put in the cloud

Ask anything.
Leak nothing.

SharpSignal reads every PDF, scan, spreadsheet and image you've got, then answers in plain English — with receipts. It all runs on your own hardware, so the smartest thing in the building never phones home.

In active development. We'll reach out the moment there's something to try.

100% air-gappedencrypted at restcited answersreads any filebuilt for FERPA / HIPAA / PCI

How it works

Drop. Ask. Done.

01

Drop in the mess

PDFs, scans, spreadsheets, Word, images, plain text. Drag it in — it's read and indexed locally.

02

Ask like a human

“What did we promise Acme?” “Which contracts renew this quarter?” No query syntax. No training.

03

Answers, with receipts

Every answer cites the exact file and line. When the files don't know, it says so — never an invented source.

Private by architecture

It never phones home.

Local-first isn't a checkbox here — it's the whole design. No document byte ever leaves the machine.

Runs on your hardware. Your machine or a small team's server — fully offline and air-gapped — nothing is sent anywhere, ever.

Encrypted at rest, with a documented deletion and retention path.

Need-to-know access scoping, enforced server-side, plus an append-only disclosure audit log.

Designed for regulated data: FERPA, HIPAA, PCI.

There are no connectors of any kind — no email, no cloud, no network ingest. Documents enter one way: you add local files. Nothing ever crosses the wire.

What you can do

Your archive just got three new jobs.

The same corpus, pointed at a different need.

recall

A memory you can interrogate

Ask across years of documents and get a straight answer. “Catch me up on Acme.” “What did we decide about the vendor?”

extract

A spreadsheet hiding in your docs

Turn a class of documents into a table: invoices into spend, contracts into renewals, resumes into a candidate list.

push

A feed that hunts for you

Standing interests fire the moment new documents land, and a morning digest brings you what's new — no checking required.

Two quiet superpowers: onboarding (new hires ask the corpus instead of your senior staff) and continuity (when someone leaves, their inbox stays answerable instead of going dark).

Reads what your office actually runs on PDF· scanned PDF (OCR)· Word· Excel / CSV· images· audio & video

How it works underneath

A few honest parts, doing one job each.

No magic under the hood — just a small set of pieces with clear roles, arranged so you can see how an answer gets made.

the backbonekeel

Always on, and the single path every model call travels. It serves those calls through a model-agnostic streaming gateway — a careful way of saying the model underneath is swappable. The apps above never have to care which one is answering today.

the visual senseeasel

The part that reads images and, before long, makes them. Vision is live now; picture-generation is coming to the shared, hosted workshop. Anything that needs to see a page or draw a scene comes here.

the deeper reachbrain

The bigger model you call on when a job genuinely needs more depth. It's an upgrade you opt into, not a dependency you're locked to — reached through the same gateway as everything else, so more depth is a routing choice.

the shared canonmemory

Persistent memory the whole workshop can read, write, and summarize — and it doesn't care which model is talking to it. Your persona and its history live here, portable across any platform, hosted or local. It's the settled record every model can draw on and none can talk its way around.

Portable persona, any platform →

Built on the corpus

The whole workshop, cooperating.

Two finished things already read from the corpus. One answers a rulebook; one remembers a world — and neither can be talked into inventing what isn't there.

RuleSage

Rules, answered with receipts.

A cited table judge for board games: ask a rules question and get the ruling with the exact passage it rests on. Built straight on the corpus, so an answer it can't source is an answer it won't fake.

See RuleSage →
“The world remembers you.” memory

Every choice and every mark you leave is written to the shared, persistent record — so when you come back, the world knows exactly who you've been, and it never quietly forgets.

“And talks back.” the backbone

Characters and places answer you in the moment, routed through the same model-agnostic gateway everything else uses — so the voice can improve underneath without the world skipping a beat.

“And can never be talked into lying.” the corpus

What's true about the world is settled and sourced. No clever phrasing coaxes a different answer, because the record — not the mood of the moment — decides what's real.

“And you can see it.” easel

The world's imagery comes from the visual side of the workshop — reading, and soon rendering, the scenes you move through. The eye is here now; the brush is on its way.

RealKeep — a world that remembers →

The obvious questions

Why not just use ChatGPT?

Because ChatGPT can't read your files until you hand them over. Every question to a cloud AI means uploading the document to someone else's computer — someone else's retention policy, someone else's subpoena, someone else's breach. SharpSignal brings the model to the files instead of sending the files to the model. Same plain-English answers; nothing to upload, so nothing to leak.

What hardware does it need?

A single modest machine — not a server room, not a cluster. The model runs resident with room to spare and streams answers faster than you can read them.

Does it ever need the internet?

No. Installing, indexing and answering are fully offline, and the models ship inside the product — nothing is fetched at runtime. Our release gate includes proving a complete run with networking disabled — there are no connectors and no network data path at all, so nothing crosses the wire.

What AI runs under the hood?

Small open-weight models that ship with the product — separate models for answering, search, OCR, page layout and reranking. Every model file is pinned by cryptographic hash, and your documents never train anything.

How do updates work if it's air-gapped?

As offline bundles you carry in. Every model weight and dependency is hash-pinned and verified before it runs, so an audit can prove exactly what executed.

Can my whole team use it — who sees what?

Yes: one machine, one organization, many people. Accounts live on the machine (no external identity provider required), what each person can see is enforced by the database itself, and every disclosure lands in an append-only audit log.

When can I try it?

It's in active development — the storage, retrieval and answering layers are built and measured; the polish is in progress. Leave your email below and we'll reach out the moment there's something to try.

Early access

Get in early.

SharpSignal is in active development. Leave your email and we'll reach out the moment there's something to try.

No spam. One note when it's ready.