PERIMETER A CompEx Company
Book an assessment

Private AI · Perception · Industrial Systems

Intelligence that never leaves the building.

We deploy language models, agents, and vision systems inside your infrastructure — your hardware, your VPC, or a sealed environment with no route to the open internet. Twenty-plus years shipping systems that run a warehouse floor, now applied to AI.

Book a Perimeter Assessment Read the case files

Already running for Benchmark Education, ADB SAFEGATE, and MassMatrix

Fixed fee · Two weeks · Ends with a priced build plan you own

Deployment mode

Perimeter deployment architecture A boundary representing your infrastructure encloses four components: an LLM runtime, a retrieval index, an agent orchestrator, and a perception pipeline. Traffic circulates between them. The route toward the public internet is blocked. Perimeter — isolated hardware, no network interface 01 LLM Runtime Open-weight models, quantized Weights loaded from local media 02 Retrieval Index Your documents, embedded and stored on your disk 03 Agent Orchestrator MCP servers to your ERP/WMS Human approval checkpoints 04 Perception Pipeline Vision, RFID, IoT telemetry Edge inference on-site No physical route out Public internet

Hardware with no network interface at all. Models and updates arrive on physical media under your chain of custody. Used where a network route is not merely discouraged but absent.

Systems shipped for

Benchmark Education ADB SAFEGATE RRW Engineering / FedEx Sony Music — Legacy Recordings MassMatrix Eli Lilly Labcorp Sodexo

01 — The problem

Most enterprise AI never leaves the pilot.

The failure mode is rarely the model. It is integration, data governance, and the absence of anyone accountable for the system after the demo. That is the part we have been doing for two decades.

95%

of organizations report no measurable P&L return from generative AI investment.

MIT NANDA, via Fortune

40%+

of agentic AI projects are forecast to be cancelled before the end of 2027.

Gartner

67%

success rate when buying from a specialist partner — versus roughly 33% for internal builds.

MIT NANDA

1 in 5

organizations have mature governance for AI agents, even as deployment accelerates.

Deloitte, 2026

We did not discover AI in 2023.

CompEx has been building perception and control systems since long before “AI” became a procurement category — a patented object image recognition engine, RFID inventory platforms, warehouse and ERP systems, airport runway signage software certified against FAA, ICAO and EASA standards.

Which means the boring parts are solved.

Data flow mapping. Role-based access. Audit trails. Change management with people who work a physical shift. Machine identity and logging are deliverables, not afterthoughts. When a model output touches an inventory record, there is a human checkpoint and a record of who approved it.

02 — Expert on demand

The expert who never leaves the floor.

Expert on Demand is Enclave, shaped into a single product — an assistant that answers only from your own manuals, SOPs, and drawings, grounded, cited, and running entirely inside your perimeter. It puts your most experienced person’s knowledge in reach of every shift, without a document ever leaving the building.

01

Source-grounded answers

Every response is generated only from the technical material you ingest, not the open internet. If it isn’t in your corpus, the assistant says so instead of guessing.

02

Zero-hallucination validation

A recursive validation layer checks each claim against its source before an answer reaches anyone. No citation, no answer.

03

Natural language & voice

Ask the way you’d ask a colleague, typed or spoken, and get an answer back the same way — hands-free when your hands are on the equipment.

04

Multi-modal

Text, photos of a damaged part, wiring diagrams, and schematics are understood together — including a live camera feed, with the answer shown back privately on a wearable display only the technician can see.

05

Workflow agents

Guided checklists and compliance verification run alongside the conversation, so an answer becomes a completed, logged task.

How an answer actually gets built

TouchscreenVoiceText chatImage uploadSchematic viewerPrivate HUD

Human and machine, working the same ticket

1

Ask

Type or speak the question, in plain language.

2

Retrieve & reason

Matches your own manuals, not the open web.

3

Verified answer

Step-by-step guidance, every claim cited.

4

Execute

The person performs the task with confidence.

5

Log

Action and sources recorded to the audit trail.

Modeled impact

30%+

faster troubleshooting and repair

25–40%

reduction in document search time

20–30%

improvement in first-time fix rate

5–10×

expertise multiplication for lean teams

Modeled, not measured: these ranges come from comparable retrieval-assisted technical-support deployments across the industry, not a named Perimeter engagement. See Proof for the results we put our name on.

Book a Perimeter Assessment See how it deploys

03 — Capabilities

Three lines. One deployment boundary.

Each line stands alone. Together they let a physical operation see, reason, and act without shipping a byte of proprietary data to a third-party inference endpoint.

Line 01

Enclave

Private AI inside your walls

Open-weight language models deployed on your hardware, in your private cloud, or on an isolated virtual machine with no network route out. No mandatory telemetry, no vendor-side prompt logging, no dependency on someone else’s uptime or terms of service.

This is the deployment shape regulated buyers, defense programs, and student-data custodians actually need — and the one most AI vendors cannot offer.

  • On-premise and air-gapped LLM deployment
  • Retrieval over your own document corpus
  • Conversational voice systems that retire legacy IVR trees
  • Model evaluation and quantization for your actual hardware
  • Data-flow maps and compliance surface documentation

Line 02

Signal

Perception and industrial systems

Our patented object image recognition technology — the engine behind Mobius — identifies items where barcodes and QR codes fail: damaged labels, no labels, moving targets, awkward angles. It has read FedEx trailers at highway speed and identified inventory through a heads-up display on a warehouse floor.

Combined with RFID, MQTT telemetry, and WMS/ERP integration, this is the layer that turns a physical operation into a queryable data source.

  • Patented object and asset image recognition
  • Vehicle and container identification at speed
  • Active and passive RFID inventory (Waventory)
  • WMS, ERP and MRP integration
  • IoT / MQTT fleet and equipment telematics

Line 03

Agency

Agentic operations, governed

Agents are only useful when wired to real systems. We build Model Context Protocol servers against your inventory, ERP, content, and communications stack, then orchestrate agents on top with explicit approval gates on any write action.

MCP has moved fast: the public server registry grew roughly eight-fold in a year and the protocol was donated to the Linux Foundation in December 2025 — it is now a safe standard to build on rather than a bet.

  • Custom MCP servers for internal systems
  • Multi-agent orchestration with human-in-the-loop gates
  • Machine identity, RBAC and immutable audit logging
  • Content, social and marketing operations agents
  • Agent evaluation harnesses and rollback paths

04 — Solutions map

Start from the problem, not the product.

Most AI vendors make you translate your problem into their taxonomy. Here it runs the other way — six things we hear from operators trying to get AI past a pilot, and exactly which line resolves each one.

“Our data can’t leave the building — but our AI vendor needs it to.”

Enclave

Sandboxed or air-gapped LLM runtime. Nothing crosses the boundary you draw.

“Barcodes and QR codes fail on the floor — damaged, unlabeled, moving too fast.”

Signal

Patented recognition reads the asset itself, not a label glued to it.

“We don’t trust an agent to act without a human checking first.”

Agency

Every write action sits behind an approval gate, with a logged decision-maker.

“Nobody here can tell us where a prompt actually went.”

Enclave

A data-flow and egress map, handed to your auditor, not just to you.

“Our WMS and ERP don’t talk to whatever agent tooling we bolt on.”

Agency

Custom MCP servers wired directly into the systems of record you already run.

“Inventory drifts all quarter and we only find out at close.”

Signal

RFID and vision events reconcile against your ERP in real time, not at month-end.

05 — Proof

Case files, with names attached.

Every figure below came from a system we built and still stand behind. The people quoted signed their names to them.

Benchmark Education Company

K–12 Publishing

A Groveport, Ohio distribution center had a 120-day shipping window carrying roughly half of annual revenue. We built the warehouse management and ERP layer on our Mobius recognition engine, including heads-up inventory identification via Google Glass on the pick floor.

“Data capture in the warehouse grew by a factor of 100.”

Tom Reycraft — Founder & CEO, Benchmark Education
100×increase in warehouse data capture
120 dayspeak window the system had to survive without failure
~50%of annual revenue moving through that window

ADB SAFEGATE

Aviation Infrastructure

Airport runway and taxiway signage design software, built to comply with FAA, ICAO, EASA and TP312 standards simultaneously. Regulatory logic encoded directly into the tool so that a non-compliant sign cannot be specified in the first place.

“CompEx delivered software our engineers rely on daily.”

Jerry Farkas — ADB SAFEGATE
4 regimesFAA, ICAO, EASA and TP312 encoded in one toolchain
Daily userelied on by their engineering team in production

RRW Engineering · FedEx

Logistics

Vehicle recognition for FedEx trailers moving at highway speed, where conventional barcode and RFID reads were unreliable. Our image recognition pipeline identified units in motion, in variable light, without stopping the vehicle.

“They solved a recognition problem others told us was not solvable.”

Randall Reed Williams — RRW Engineering
At speedidentification without halting the asset
Patentedrecognition method underpinning the pipeline

MassMatrix

Life Sciences

A secure biological data platform on AWS for proteomics research, built around controlled access to sensitive research data — the same governance discipline that now underpins our private-model deployments.

“A platform we could trust with sensitive research data.”

Dave Ditmars — MassMatrix
Cloud-nativeAWS architecture with scoped data access
Research-gradecontrols over sensitive biological datasets

06 — Engagements

We publish our prices. Almost nobody does.

You should be able to size this before you take a call. Every engagement starts with the same two-week assessment so nobody commits to a build before the architecture is on paper.

Start here

Perimeter Assessment

$12,500

Fixed fee · 2 weeks

  • Reference architecture for private deployment
  • Data-flow and egress map
  • Compliance surface review
  • Hardware sizing and cost model
  • Fixed-price build plan — yours to keep

Credited in full against a build started within 90 days.

Line 01

Enclave Deployment

from$85,000

8–14 weeks

  • Private or air-gapped model runtime
  • Retrieval over your corpus
  • Access control and audit logging
  • Runbook and staff training

Scope and price fixed by the assessment before work begins.

Lines 02 & 03

Signal & Agency Builds

from$60,000

6–16 weeks

  • Vision, RFID or telemetry pipeline
  • MCP servers against your systems
  • Agent orchestration with approval gates
  • Integration into WMS / ERP of record

Hardware quoted separately at cost plus integration.

Ongoing

Embedded Team

$10–30k / month

Rolling, 90-day notice

  • Named senior engineers, reserved capacity
  • Model and agent operations
  • Quarterly architecture review
  • Roadmap and internal enablement

12-month commitments discounted 10% billed quarterly in advance.

What we don't claim

We are not SOC 2 or ISO 27001 certified, and we will not imply otherwise. What we do instead: build to the control patterns those frameworks describe, deploy inside your compliance boundary so your existing certifications and audits cover the workload, and hand over the documentation your auditor will ask for.

We also work in defense and public sector contexts where we do not name the program, the agency, or the contract. If that matters to your evaluation, ask us directly and we will tell you what we can under the constraints we have.

07 — The alternatives

Three ways to buy this. Honestly compared.

We are not the right answer for every organization. Here is where each option actually wins, including the two that are not us.

Comparison of API-first AI vendors, large systems integrators, and Perimeter
Dimension Option A

API-first AI vendor

 Option B

Big 4 / global SI

 Option C

Perimeter
Where your data goes Their inference endpoint, under their terms of service, with prompt retention you negotiate rather than control. Usually a hyperscaler tenancy they configure. Better, but the boundary is still someone else’s. Nowhere. Weights, index and logs sit on hardware inside your compliance boundary.
Time to first production system Days to a demo. The gap is production: only about 5% of custom enterprise tools reach it. Six to eighteen months, gated on discovery phases and steering committees. Two-week assessment, then 8–14 weeks to a running deployment with a runbook.
First-year cost, order of magnitude Low entry, per-seat or per-resolution metering that scales with success rather than value. €150K–500K for a mid-size programme, before internal cost. $12,500 assessment, $60K–$85K+ build, optional $10–30K per month embedded.
Who does the integration You do. The vendor ships an API and a solutions engineer. A rotating bench, often staffed junior once the pitch team leaves. The same senior engineers who scoped it. No handoff, no bench rotation.
Physical-world systems Out of scope. Vision, RFID and WMS integration are your problem. Subcontracted, usually to a specialist like us. In scope. Patented recognition, RFID and ERP integration are the original business.
Pick this when Your data is not sensitive, the use case is generic, and speed beats control. You need global rollout, change management across thousands of staff, and audit-committee cover. Data cannot leave, the workflow touches physical operations, and you want one accountable team.

Cost bands for options A and B are market benchmarks, not quotes. Production and pilot-failure figures from the MIT NANDA study reported by Fortune. Our own figures are the ones published in the engagements table above.

08 — Sectors

Where the boundary actually matters.

Private deployment is not a philosophical preference in these industries. It is a procurement requirement, a statutory one, or both.

Education & Publishing K–12

More than thirty states have now codified rules on AI in schools, and district procurement is following. Idaho SB 1227 requires aligned procurement policy; California's AB 2885 and AB 1008 pull AI-processed student data squarely inside existing privacy law. A model running on district or publisher infrastructure sidesteps the hardest questions in the RFP.

Student data residencyHuman-in-the-loop gradingCurriculum retrievalState AI policy alignment

Defense & Public Sector Restricted

Localized language models on isolated hardware, with augmented-reality interfaces for field use. Built for environments where there is no network to call out to and no acceptable circumstance in which data leaves the enclave. We describe the capability; we do not describe the customer.

Air-gapped inferenceEdge hardwareAR interfacesNo-egress architecture

Logistics & Industrial Floor-level

Warehousing, distribution, fleet and agricultural equipment. Recognition and RFID at the point of physical work, tied back into the WMS or ERP that runs the business. This is the environment our patented recognition technology was built for and proven in.

WMS & ERP integrationRFID inventoryVision at speedEquipment telematics

Life Sciences & Regulated Data Controlled

Research and clinical data that cannot be posted to a public endpoint. Retrieval and analysis over proprietary corpora, running inside the customer's existing validated environment so that qualification and audit trails remain intact.

Proprietary corpus retrievalScoped accessValidated environmentsAudit trails

09 — Perimeter Labs

What we are building next.

Stated plainly: these are in active development, not shipping products. We list them because they show where the three capability lines are converging — and because early partners help shape them.

Lectern — concept film, not a shipping avatar

In development

Lectern

A three-dimensional avatar instructor for K–12 and training environments. Voice-driven, curriculum-grounded through retrieval over the publisher’s own materials, and able to run against a private model so no student utterance leaves the district. Aimed at the gap between static video and a live teacher.

In development

Voice Enclave

Fully local speech-to-text, reasoning and speech synthesis on a single on-premise appliance. Conversational systems for call handling, field operations and accessibility that work with the network cable unplugged.

In development

Floor Agent

Perception and agent orchestration merged: vision and RFID events trigger agents that reconcile inventory, flag exceptions and draft the corrective action — then wait for a human to approve it before anything is written.

In development

Voicekey

Passwordless authentication from a spoken passphrase. Enrollment and matching happen entirely on-device — no voiceprint database, nothing to breach, nothing to phish. Works with the network cable unplugged, same as Voice Enclave.

10 — Questions

What people ask before they call.

Straight answers to what usually comes up when a company starts looking at private AI integration.

How do we integrate AI into our company without sending data to the cloud?

Perimeter deploys the model, the retrieval index, and the agent layer on hardware you control — your own servers, your private cloud tenancy, or a fully air-gapped machine with no network interface at all. Nothing about a prompt, a document, or a response reaches a third-party inference endpoint. See deployment modes for the three shapes this takes.

What is air-gapped AI, and does our company need it?

Air-gapped AI runs on hardware with no network route to the outside world — models and updates arrive on physical media under your chain of custody. It's the right fit when a network route isn't just discouraged but contractually or statutorily absent, which comes up most often in defense, life sciences, and student-data contexts.

Can an AI assistant answer from our own manuals instead of the open internet?

Yes — that's Expert on Demand. It's a retrieval-augmented assistant that only answers from the technical material you ingest: manuals, SOPs, schematics, drawings. If the answer isn't in your corpus, it says so instead of guessing, and every claim is cited back to a source before it reaches anyone.

How long does private AI integration take to deploy?

Every engagement starts with a two-week, fixed-fee Perimeter Assessment that ends in a priced build plan you own. The build itself typically runs 6–16 weeks depending on scope — see Engagements for the full breakdown.

What does private AI integration cost?

Perimeter publishes its pricing rather than gating it behind a sales call: a $12,500 two-week assessment (credited in full against a build started within 90 days), Enclave deployments from $85,000, Signal & Agency builds from $60,000, and embedded engineering teams from $10,000–$30,000 a month.

How is this different from using ChatGPT or another cloud AI tool at work?

A cloud AI vendor processes your prompts on their infrastructure, under their terms of service, with retention you negotiate rather than control. Perimeter runs a comparable class of model entirely inside your compliance boundary — your hardware or your private cloud — so you get similar capability without a document ever leaving your perimeter. See the full comparison.

Do you support voice or hands-free interfaces for field and technical work?

Yes. Natural language and voice run alongside touchscreen, text chat, image upload, and a private heads-up display, so a technician can ask a question hands-free and get the answer back without looking away from the equipment.

11 — Contact

Start with the assessment.

Tell us what you are trying to keep inside your walls and what it currently touches. If a two-week assessment is not the right first step, we will say so — we would rather decline than run a project that cannot ship.

Office

470 Olde Worthington Road, Suite 200
Westerville, Ohio 43082

Good fit

Organizations with a real constraint on where data can live, an operation that already runs on systems of record, and a decision-maker in the room.

Received. We reply within one business day.

Three fields, one reply from an engineer — not a sequence of marketing emails.

Fixed fee · Two weeks · No pressure Book Assessment