Your project's AI, running on your own floor.
A complete system: the machine, the models and the construction.live harness, delivered as one thing. Your drawings, contracts, daily logs and voice notes are read on a box in your building — never sent to a model vendor, never metered by the token.
- 128 GB
- Unified memory
- 32 GB
- Dedicated GDDR6
- 191
- TFLOPS at FP16
- 0
- Bytes sent to a model vendor
Built for inference
128 GB of unified memory holds a long-context model resident, and a dedicated 32 GB GDDR6 card carries the decode. Big enough for a full drawing set in one prompt, fast enough that a super isn't waiting on it.
Ready in an afternoon
It arrives with the engine, the models and the construction.live harness already on it. Power, network, pair, and your projects are indexing before the crew goes home.
Private by default
Every drawing, contract, daily log and voice note is read on the box in your building. Nothing is sent to a model vendor, because there is no model vendor in the path.
Three ways to put it in the building.
Every configuration ships the whole platform. What changes is where the box sits, how much of it there is, and how tightly it's sealed off from everything else.
Site Box
Jobsite trailer or a single office
One box, one job.
The base unit, sized to sit under a desk in the trailer. It runs the whole platform for the crew on that job — voice notes, daily logs, drawing revisions, submittals — without a line out to anything.
- Units
- 1
- Unified memory
- 128 GB
- Dedicated VRAM
- 32 GB
- Draw
- ~500 W sustained
- Full construction.live platform, on-box
- DeepSeek V4 and Qwen3.8-27B pre-tuned
- Unlimited seats on that project
- Wi-Fi 7, 5 GbE, runs off a standard 15 A circuit
Office Cluster
A contractor running a portfolio
Unlimited use, flat cost.
Four units racked together in your own server room, serving every project and every seat in the company. You stop paying per token and start paying for a machine, so nobody in the office has a reason to ration what they ask it.
- Units
- 4, clustered
- Unified memory
- 512 GB aggregate
- Dedicated VRAM
- 128 GB aggregate
- Draw
- ~2 kW sustained
- Everything in Site Box, across every live project
- Unlimited seats, unlimited usage, no per-token metering
- OpenAI- and Anthropic-compatible endpoints for your own tools
- Procore, Autodesk, Bluebeam and email connectors run on-box
Government Enclave
Public owners, defence and infrastructure
Completely self-hosted. No egress.
The same platform, delivered air-gapped. It is installed inside your perimeter, the weights are on your disks, and the box has no route to the public internet — not a restricted one, none. Updates arrive the way anything else arrives in a secure facility: on media you inspect first.
- Units
- 4 to 16, rack-mounted
- Unified memory
- Up to 2 TB aggregate
- Dedicated VRAM
- Up to 512 GB aggregate
- Egress
- None
- Air-gapped install, offline model and platform updates
- Data residency by construction: nothing leaves the facility
- Full audit log of every prompt, retrieval and generation
- SSO, role-based access and per-project compartments
Quoted per deployment rather than off a list, because unit count, siting and integration scope all move the number. Tell us the shape of yours and we'll come back with a figure.
Open weights, on your disks.
The box arrives with two models already tuned for construction work and the harness already routing between them. No licence server to reach, no model that can be deprecated out from under a contract.
DeepSeek V4
Long-context reasoningCarries the work that needs the whole file in view at once: reading a drawing set against a schedule, tracing a change order back through the revisions that caused it, reconciling an invoice with what the daily logs say happened.
Qwen3.8-27B
High-throughput field workThe model that answers while a super is still standing in the stairwell. Voice note transcription and filing, daily log drafting, submittal classification, search across everything on the job.
A model on its own doesn't run a project.
What makes it a system is the construction.live harness: the agents that read what arrives, file it against the right job, link it to the revision behind it and follow up on what's owed. That ships on the box too.
It reads what arrives
Email threads, attachments, drawings, voice notes and photos land on the box and are parsed there. The agents that do the reading never leave the machine.
It files against the right job
Every artefact is matched to the project, the drawing revision and the schedule behind it, so the connected record builds itself as the job runs.
It answers on your endpoints
The same OpenAI- and Anthropic-compatible API your own tools already speak. Point an internal script, a BI tool or a coding agent at it and it just works.
For government, completely local.
Public infrastructure work carries records that can't be handed to a third party: sealed bids, security-sensitive drawings, procurement correspondence, claims headed for litigation. The Government Enclave configuration answers that the only way it can actually be answered — by putting the whole system inside your perimeter.
Nothing leaves the building
In the air-gapped configuration the box has no route off your network. There is no telemetry channel, no phone-home, no vendor-side inference. Privacy here is a property of the wiring, not a clause in a contract.
The weights are yours
Open-weight models on your own disks. No licence server to reach, no deprecation notice that retires the model your procurement was written around, no vendor able to change the terms after award.
Auditable end to end
Every prompt, retrieval and generation is logged on-box with the record it touched, so an FOI request, an inquiry or an internal review can be answered from your own logs.
Priced as a capital asset
One machine, one purchase order, unlimited use. Nobody has to forecast token spend, and departmental usage never becomes a budget conversation.
An aluminium chassis built for sustained load, not benchmarks.
Per unit. The Office Cluster and Government Enclave configurations multiply these.
- Processor
- AMD Ryzen AI MAX+ 395
- Integrated GPU
- 40 cores, up to 60 TFLOPS
- Dedicated GPU
- AMD Radeon AI PRO R9700, 64 CUs
- GPU performance
- Up to 191 TFLOPS at FP16
- Unified memory
- 128 GB LPDDR5X-8000
- Unified bandwidth
- 256 GB/s
- Video memory
- 32 GB GDDR6
- Video bandwidth
- 640 GB/s
- Capacity
- 2 TB NVMe 1.4
- Drive
- KingSpec XG7000
- Sequential read
- Up to 7,400 MB/s
- Sequential write
- Up to 6,600 MB/s
- Supply
- 1,000 W Corsair SF1000, 80 PLUS Platinum
- Sustained load
- ~500 W
- Idle
- ~40 W
- Circuit
- Standard 15 A office outlet
- Dimensions
- 340 × 110 × 320 mm (13.4 × 4.3 × 12.6 in)
- Volume
- 11.97 L (~730 in³)
- Weight
- 6 kg (13.2 lb)
- Build
- Machined aluminium, air-cooled for low noise
- Networking
- 5 GbE Ethernet, Wi-Fi 7, Bluetooth
- Data ports
- 2× USB4, 2× USB-A
- Display
- HDMI 2.1, 2× DisplayPort 2.1, 4 GPU outputs
- Air-gap option
- Radios disabled in firmware on request
- Operating system
- Ubuntu
- API
- OpenAI- and Anthropic-compatible
- Pre-installed
- Inference engine, models, construction.live harness
- Pre-ship testing
- 72-hour burn-in
Built to order, in limited numbers.
Tell us the deployment
Trailer, server room or secure facility, and how many projects and people it has to carry.
We size and quote it
Unit count, siting and integration scope, with the connectors you already run costed in.
We build, burn in and install
72 hours of load testing before it ships, then installed and indexing your projects on site.
- 1 year
- Parts and labour
- 72 hours
- Burn-in before it ships
- On site
- Installed and handed over
Before you ask us for a number.
No. The box runs the whole platform. Cloud and on-premise are the same product with the weights in a different room, and projects can move between them.
Tell us where the box goes.
Trailer, server room or secure facility. We'll size it, quote it and tell you honestly if the cloud version would serve you better.