Perimeter is an AI appliance that reads your company's documents and answers with sources — on a box in your server room. It makes zero external calls. Pull the network cable and it keeps working.
Live in 30 days · annual licence · hardware bought in the EU
Matching supplier catalogues to line items, digging through contracts and standards — skilled people spend most of their week finding things, not deciding things.
Matching in minutes, answers in seconds — each one with a link to the source document, so you can check it.
Bid pricing, supplier terms, client files. Sending them to a US-jurisdiction cloud is a commercial risk before it is a legal one — and for many of your customers, a dealbreaker.
Nothing ever leaves. The appliance works without the internet — and you can verify that yourself, live.
Open-source stacks are real, but they are an engineering project: a server, pipelines, evaluation, upkeep. The build-vs-buy line sits around three dedicated engineers.
A box arrives, we set it up on your corpus, your team is working in 30 days. No platform team required.
The big platforms start around $100k a year, per-seat, and are scoped for Fortune 500 procurement — not for a 120-person company with one estimating office.
One box, one annual licence, sized for mid-market.
Perimeter is storage for meaning, not files. Ordinary search matches exact words; Perimeter finds things by what they mean — so "steel pipe 57×3.5" and its eleven other spellings resolve to the same item. It learns only what you load into it, inside your walls.
Catalogues, BoQs, contracts, standards, matters — loaded locally, never uploaded anywhere.
A GX10-class box in your server room. Open-weight models. Zero external calls.
Every answer links to the document and place it came from. Trust, but verify — in one click.
Fair — and common. The difference is not the AI model. It is where your documents live, and who can reach them.
| Perimeter | Cloud AI | DIY open-source | Enterprise suite | |
|---|---|---|---|---|
| Where your data lives | Your server room | Provider's regions | Yours — once built | Vendor cloud / heavy on-prem |
| Time to working | 30 days | Instant — with the risk | Months | 6–12 month rollout |
| Who you need to hire | No one | No one | ~3 ML engineers | A project team |
| Cost shape | Setup + annual licence | Per seat, forever | Salaries, forever | $100k+/year floors |
| Verifiable in person | Yes — pull the cable | No | If you built it right | Rarely |
We fix a reference sample of your real documents and agree an acceptance threshold in writing — before we start. Accuracy and hours saved are measured on your corpus, not claimed from a slide.
The box arrives, we load your corpus, train your team in two sessions, and run the load test for your real number of users. If the pilot misses the gate, no licence invoice.
Annual licence, support at 20% — including model updates. Open models improve every few months; with support, your box is always on the current one.
Match supplier catalogues to bill-of-quantities line items in minutes instead of days. Quantification eats 50–80% of an estimator's time — and bid pricing can't go to a cloud. This is the problem Perimeter was built on.
The construction page Law firmsMatters, contracts, memos — searchable in seconds, with citations. Client files never leave the office.
The legal page Your industryBig catalogues, confidential terms, document-heavy teams. If that sounds like you, show us your documents.
Talk to usPerimeter processes everything inside your perimeter, so there is no cross-border transfer to assess. Hardware is bought in the EU in your name; models are open-weight with public provenance; the contract is with a Dutch BV. Architecture review and code audit on request.
One team, one corpus. The honest way to try sovereign AI.
The full appliance for a 50–250-person company: matching, search, training, load-tested for your users.
Heavier hardware, more users, deeper integration, priority SLA.
Licence is annual, paid upfront. Support is 20% and includes model updates. Hardware at cost +15–20%, bought in the EU under your name. Early customers can join as a founding reference: a setup discount in exchange for a measured, public case study. Full pricing
That is exactly what the pilot measures. We agree a reference sample and an acceptance threshold in writing before we start; if the pilot misses the gate, you don't pay for the licence — only the pilot work.
Honestly: it depends on your documents and usage, so we load-test your real profile during the pilot instead of promising a number. One box comfortably serves an interactive team; for heavier loads there is a faster model configuration and a heavier hardware class.
No — model updates are included in annual support. The model is a slot, not the product: when a better open-weight model ships, we test it on your corpus and upgrade your box.
HumanOps B.V., a Dutch company based in Amsterdam. The engineering team has Russian roots; the company you contract with is Dutch. Your data never leaves your building, the hardware is bought in the EU, the models are open-weight, and we offer an architecture and code review on request. We'd rather you ask than wonder.
No. Perimeter reads your documents and answers questions about them. Your ERP, estimating suite or legal database stays; Perimeter removes the manual searching and matching around them.
The box is yours — you bought it. The corpus and the index live on it, inside your walls. We keep no copies; there is nothing to give back.
Twenty minutes, no slides. Bring a page of your catalogue or a contract — we'll show the matching and the sources live.