Niobe
Cited procedural intelligence · at the point of work
NIOBE
Private intelligence system · keyed to one operator

Your technician has the manual in their ear. Even three decks down.

Niobe answers from your own technical library — cited to the page, spoken hands-free, running on hardware you control. No signal required. No upload required.

Scope a deploymentRead the measurements

TRL 4 · demonstrator measured 22 Aug 2026 · CMMC Level 2 self-assessment in preparation, no third-party certification claimed

Niobe Edge · air-gappedNo network

“How do you treat heat exhaustion?” — ask it above and the answer comes back from the manual itself, with the page. Try something it does not cover, too: refusing is the part worth testing.

Your question is not stored. We keep a daily count of how often this box answers, refuses, or finds only related material — not what was asked. Rate limiting uses a token that rotates every hour, never your address.

Two publicly released U.S. Navy manuals are loaded: NAVEDTRA 14150 (Machinist’s Mate 1 & C) and NAVEDTRA 14295 (Hospital Corpsman).

2manuals loaded
0uncited claims
—coverage — ask to measure
Government & defense

The network is denied and the manual is 400 pages.

Cited procedure steps at the point of work, grounded in your own unclassified library, running fully air-gapped. Entity, registration and cybersecurity posture stated plainly — including what is not yet in place.

Entity & readiness → Solo & small business

You are the whole back office.

The same engine, pointed at the work that repeats — inbox, follow-ups, quotes, paperwork, the opportunities you never get time to chase. You can run it yourself today.

What it handles →
The problem, in three rooms

The knowledge exists. It's never where the hands are.

A shipboard technician, an elevator mechanic, and a firm whose files can't leave the building have the same problem in different clothes: the right procedure, at the point of work, without stopping to go find it.

Engine room · deck 3

No signal, no manual.

The tech climbs back up, finds the binder, climbs back down. Twice. The repair takes four hours instead of one, and the fix depends on what he remembered on the ladder.

Basement riser · commercial

Hands full, phone useless.

Zero bars, gloves on, a schematic he'd need three hands to hold. So he guesses, or he calls the one person in the company who knows — if they pick up.

Practice · files under seal

Can't upload, so can't use.

The firm has twenty years of precedent on a server. Every useful AI tool starts with “upload your documents,” and that sentence ends the conversation.

Same engine. Three sets of stakes.
01 · Defense & naval

Where the network is denied.

  • Walk a repair with the manual in your ear, steps on your visor
  • Grounded in your own unclassified NAVSEA / Army library
  • 100% air-gapped — keeps working when every link is gone
  • Stale readings labelled, never drawn as live
02 · Commercial trades

Where your hands are full.

  • Say the equipment, see the schematic — step by step
  • Voice built for gloves, masks, and loud rooms
  • Electrical, plumbing, elevators, industrial installs
  • Works with zero bars — basements, plants, vessels
03 · Rooms with rules

Where files can't leave.

  • Legal, medical, accounting — read on your own hardware
  • There is no upload; custody is a build fact
  • Tamper-evident activity log — content never recorded
  • Not a policy you trust — architecture you can verify
05▶Custody

Data That Can't Leave

Your files, read on your own hardware. There is no upload.

06▶Edge

Works Where Signal Dies

Job sites, vessels, basements — Niobe Edge stays on.

Live capture of the Niobe heads-up display: an answer cited to NAVEDTRA 14295, a public U.S. Navy training manual, beside a twelve-feed tactical picture with an escalating shipboard alert.

Recorded from the working system. The answer on the left is cited to NAVEDTRA 14295 — a real, publicly released U.S. Navy training manual — while twelve live readings refresh on the right. Simulated sensors, real software, zero network.

Generic AI works like a stranger who just met you. It forgets you by the next session.

I · The Brain

It learns you.

A deep interview maps how you think, decide, and move — then it indexes your technical library — manuals, standards, schematics, procedures — however large it happens to be. It sharpens with every correction and never makes you re-explain yourself again.

↳ a mind that's yours
II · The Workforce

It works while you sleep.

Then it goes to work, hands-free — hunts your next contract, clears the admin, overlays live procedure steps on a heads-up display, and ships the work in your voice. You wake to things done, not a to-do list.

↳ hands-free execution · 24/7
Meet Niobe
Every other AI knows your last prompt.
NIOBE knows you.
◐01▶
What Niobe Is
An AI employee that runs your business
✦02▶
How It Works
Set up in one sitting, not a project
⊕03
Keyed to You
Your library, your rules, your corrections
∿04▶
Does the Work
Real work done — nothing sent without your ok
◈05▶
Your Back Office
Inbox, calendar, quotes, and invoices
Measured, not asserted

Everything above is a claim. This part is a number.

One demonstrator run, reproducible from the source repository, with the conditions stated rather than footnoted.

Automated tests
270 pass · 27.6s

The suite runs with no model present, so a green result is the architecture, not a lucky generation.

Tactical path p95
0.541 ms

Sensor → bus → fusion → emit, against a 30 ms budget. 34 samples, zero breaches.

Adversarial screening
49 / 49 refused

And 19/19 legitimate prompts passed. A system that refuses everything is broken, not safe — the second number is what separates them.

Audit integrity
47 frames, chain intact

Hash-chained HUD emissions, zero dropped sensor readings.

What this run does not show. It executed on a development workstation using recorded sensor replay and stubbed inference. It demonstrates that the architecture holds its latency budget and that the refusal controls fire. It does not demonstrate on-device performance — latency and endurance with a quantized model on target headset hardware are unmeasured, and establishing them is Phase I work. We assess the system at TRL 4. We hold no third-party assessment and claim none. And we have no completed customer deployment — you would find that out in diligence, so you should find it out here. See the full run →

How the two paths stay out of each other's way
How the tactical and advisory paths share one display Two independent chains converge on a single heads-up display. The upper tactical chain runs sensor bus to fusion to emit on a fixed ten-millisecond cadence against a thirty-millisecond budget, never waiting on a model. The lower advisory chain runs voice input to retrieval over your library to a local model, measured in seconds. Neither chain can block the other; they share only the glass. HARD REAL-TIME · 10 MS CADENCE · 30 MS BUDGET Sensor bus12 feeds Fusionstate + staleness Emithash-chained HUD one display SOFT ADVISORY · MEASURED IN SECONDS Voice ingloves, masks Retrievalyour library Local modelon your metal
Tactical

It never waits on a model.

Sensor state reaches the display on a fixed cadence against a 30 ms budget, whether or not anyone is asking a question. Readings that go stale are labelled stale — never drawn as live.

↳ hard real-time · shares only the glass
Advisory

It can only cite, never invent.

Answers are retrieved from your ingested library and returned with the document and the page. If your library doesn't cover it, Niobe says so instead of guessing. The obvious design — one loop that thinks, then draws — freezes the operator's picture for the length of every inference. We built that version and rejected it.

↳ grounded · refuses when uncovered
Entity & registration, stated plainly

Including what is not yet in place. Every value here is generated from a single status file and verified against it before deploy. The entity was formed on 20 August 2026; the engineering predates it. We would rather you read both facts here than discover them later.

Legal name
Hey Niobe LLC
Entity type
Texas LLC
TX SOS file
806756936
Formed
2026-08-20
SAM.gov
Registration in progress
UEI
Pending validation
CAGE
Pending — DLA
CMMC
L2 self-assessment, in preparation
Third-party assessment
None claimed
Primary NAICS
541715
Business size
Small business
Managing member
Razaak R. Raji

Full readiness detail →

Where it runs

The voice in your ear today. The face in your glasses tomorrow.

NIOBE runs on what you already own — your phone and your earbuds — and on local edge hardware when the job leaves coverage. Heads-up display streaming for AR smart glasses is built in, not bolted on. See Niobe Edge →

PhoneEarbudsAir-gapped EdgeAR Glasses · HUD · in build
From the founder
“Niobe ran my own business for months before I offered it to anyone. It caught a duplicate charge and a grant deadline I would have missed — that’s the day I knew it was real. I built it for myself first. Now it learns you.”
— Razaak “Zack” Raji, founder of Niobe
Pricing

Priced to the work.

Per operator, scoped to your library and the size of your crew.

A paid 30-day pilot comes first. We agree in writing what has to be true for it to have worked. If it does not hit those criteria, there is no deployment for us to sell you.

Book a scope call See what we measured
The floor

We would rather lose the sale. If your library does not cover the work, we say so on the scoping call — before anyone signs anything. That is a cheaper answer for both of us than a deployment that stalls.

One conversation · 45 minutes · it can end in a no
NIOBE

Bring one repair or one procedure that keeps going wrong. We will tell you honestly whether your library covers it — including if it doesn't.

Book a scope callRead the measurements
01✕