documents
Reads what your team reads
Scanned PDFs, phone photos, spreadsheets, emails and drawings. The agent takes what arrives, in whatever format it arrives.
AI agents that do real work, on infrastructure we run.
We build custom AI agents for the document-heavy work your team still does by hand: reading quotes, checking drawings, filling schedules, pricing bids. Each one is built and scored against your own past work before it touches a live job. We also deploy and operate dedicated NVIDIA DGX B300 systems.
AGENTS
CUSTOM_BUILT · SCORED_ON_YOUR_PAST_WORK
IN_PILOT
TENDERING_AGENT
INFRASTRUCTURE
NVIDIA_DGX_B300 · DEDICATED
AI agents, built for your work.
We start from a job your team already does and the results they've already produced, then build an agent that does the reading and checking for them. A person still makes every decision that matters.
sources
Every answer shows its source
Each figure says which document and page it came from and how sure the agent is, so checking it takes seconds.
guards
Rules in code, not in prompts
When a model gets something wrong, we fix it with a tested check in code. A model can't be relied on to follow an instruction every time.
people
Your team decides
The agent proposes and flags. Your people accept or reject, and every decision is logged with who made it and when.
measured
Scored against your real results
Before an agent goes near live work, we measure it against what your team actually submitted in the past.
models
Your choice of model
Run on a commercial model API, or on open models hosted on dedicated hardware so your documents never reach a third party.
Example: a tendering agent.
Built with an electrical contractor whose estimators price large building tenders. It handles the reading and cross-checking that used to take days, and it now runs beside live tenders without changing what the firm submits.
quotes Built
Supplier quotes
Reads every supplier quote, whether it's a scan, a photo, a spreadsheet, an email or prices handwritten on a printout. Matches each price to the firm's price list and flags big moves from earlier revisions and past tenders.
takeoff Built
Quantities from drawings
Reads quantities from the consultant's drawings and checks them against a second count on the same drawing. Where the drawing contradicts itself, it shows the excerpts side by side.
pricing Built
Pricing the bid
Prices the bill of quantities from this tender's quotes, the firm's past bids and its rate library. Every line says where its figure came from, and lines it can't price are listed with what's missing.
compliance Built
Technical schedules
Fills the tenderer's technical offer from the suppliers' own data returns and checks approved makes. Every answer cites its source, and the rows that need an estimator are marked.
qa Built
Pre-submission check
A rule-based check of the priced bill before it goes out. No model involved, so the same input always gives the same answer.
pilot In pilot
Running beside live tenders
Estimators use it alongside their usual workbook. After each submission, every part is scored against what was actually sent, and the tender joins the history the next bid learns from.
Tendering isn't the only fit. Any job where people read a pile of documents, match them against a list and fill in a schedule is a candidate. Tell us about yours.
How we build an agent.
No big-bang rollout. The agent earns trust on your past work first, then on live work beside your team.
01 · map
Learn the job
We sit with the people who do the work and collect a few past jobs, with the documents that went in and the result that went out.
02 · build
Build and score
We build the agent against those past jobs and measure it on them, so you see how often it's right before anyone relies on it.
03 · run
Run it beside your team
It works alongside a live job without changing what you deliver, gets scored after each one, and takes over more as the numbers hold up.
Private AI infrastructure.
We own and operate NVIDIA DGX B300 systems: eight Blackwell Ultra GPUs per node on NVLink. Our current node is on a long-term dedicated lease. Talk to us about the next one.
dedicated On request
Dedicated nodes
A whole DGX B300, bare metal, with your own operating system and network. Fixed monthly rate, long-term.
operations Included
We run the hardware
Data centre, power, cross-connects, firmware and out-of-band management are ours to look after. You get console and power control over a scoped, logged account.
inference On request
Private model hosting
Open models like Qwen, Llama or DeepSeek served on dedicated hardware behind an OpenAI-compatible endpoint, so agents can run without your data leaving it.
sandbox Early access
Agent sandboxes
Isolated, disposable environments for running an agent's code, with secrets the agent never sees. How they work.
confidential Coming soon
Confidential computing
Intel TDX with GPU confidential-computing mode, so memory is encrypted even from us. Being commissioned on this hardware, not enabled yet.
storage Included
Encrypted storage + erasure certificate
Volumes are encrypted at rest (LUKS2) and cryptographically erased at hand-back, with a certificate you can verify.
| A DGX B300 node | Specification |
|---|---|
| GPUs | 8× NVIDIA B300 SXM6, 288 GB HBM3e each (2.3 TB total), NVLink between all eight |
| CPU and memory | 2× Intel Xeon 6776P, 2 TB DDR5 |
| Networking | 8× ConnectX-8 SuperNIC, 2× BlueField-3 DPU |
| Storage | Local NVMe |
Security you can check, stated plainly.
Agents keep an append-only log of what they proposed and what your people decided. You choose where the model runs: a commercial API, or dedicated hardware that we own and run, in one data centre. Volumes are encrypted at rest and crypto-erased at hand-back, with a certificate. Confidential computing is being commissioned, not enabled, and we're not certified against any compliance framework yet. We'll always tell you exactly what's on.
terrabox
agents + DGX B300
AGENT_ANSWERSSOURCED · HUMAN_REVIEWED
DECISION_LOGAPPEND-ONLY · WHO + WHEN
MODEL_HOSTINGAPI OR DEDICATED · YOUR CHOICE
NODE_VOLUMESLUKS2 AES-XTS · CRYPTO-ERASE
CONFIDENTIAL_TIERCOMMISSIONING · NOT ENABLED
$
./scripts/tbxcert.py verify TBX-CE-2026-0001