Local BI.
Ask your
numbers.

Business intelligence that runs locally: an AI analyst trained on your company's data, on your racks or our EU servers, so your data stays yours. Ask a question, get the number and where it came from.

ON SITE QUERIES 0 LEFT THE BUILDING 0 B

ASK IT ANYTHING YOUR DATA CAN ANSWER

AI ANALYSTEXAMPLE

YOU ASK

IT WRITES THE QUERY · AND SHOWS IT

SELECT region,
       SUM(revenue - cogs) / SUM(revenue) AS gross_margin
FROM   erp.sales_lines
WHERE  quarter = '2026-Q3'
GROUP BY region ORDER BY gross_margin DESC;

THE ANSWER · WITH ITS SOURCE

NORTH41.2%

WEST38.7%

SOUTH33.9%

EAST29.4%

erp.sales_lines · 14,212 rows · “gross margin” as your finance team defines it

Documents too: contracts, specs, policies and past reports, searchable in plain language, every answer linking to the page it came from. We have shipped this before: ArhiTech runs a self-hosted model in the EU over permit drawings and documents.

HOW IT GETS BUILT

YOUR DATAYOUR TEAM

SOURCES3

EVAL--

GPUS--

DATA OUT0 B

CONNECT

We wire the AI into what you already run: ERP, CRM, databases, documents. One connector layer, read-only by default, inside your network.

TRAIN

An open model, adapted to your data and your questions with a small LoRA adapter, and scored on an eval set built from your own past reports. “It works” becomes a number.

SERVE

Sized to fit one or two GPUs in your racks, or a dedicated server in the EU. Quantized and batched the way our inference work does it, so it is cheap to run.

LEARN

Every correction your team makes is training data. A monthly retrain, gated by the same evals, so it gets better instead of drifting.

WHERE IT RUNS

YOUR RACKS

ON-PREM · AIR-GAPPED IF YOU NEED IT

We size the hardware to the model (usually one or two GPUs, not a data centre), install it, and run it inside your network. Already have GPUs? We use those.

Hardware at cost, or none if you have it
No outbound connection required
Your IT owns the box

OUR EU SERVERS

DEDICATED · HOSTED IN THE EU

A dedicated server in an EU data centre, used by you and nobody else. No hardware to buy, nothing shared with other clients, and your data stays under EU law.

Dedicated, never shared
EU jurisdiction, GDPR by design
Move it on-prem later if you want

ENGAGEMENTS · PRIVATE AI

SETUP

4–8 WEEKS · FROM €8,000

Connectors into your systems, a model adapted to your data, an eval set built from your own questions, and deployment on your racks or our EU servers. Ends when it passes the evals.

Read-only connectors, inside your network
Evals you keep and can rerun
Every answer cites its source

VLLMOLLAMAQDRANTPOSTGRESYOUR ERP

RUN & LEARN

MONTHLY · FROM €1,200/MO

We keep it running and keep it improving: monitoring, a monthly retrain from your team's corrections, model upgrades as better open models ship, and a human on the other end.

Monthly retrain, gated by your evals
Upgrades to newer open models
Cancel with 30 days' notice

YOUR HARDWAREYOUR DATAYOUR MODEL

SECURITY & ACCESS

The model, the data and every answer stay on your hardware or your dedicated EU server.
Connectors are read-only unless you decide otherwise, and respect your existing permissions.
Our access is least-privilege, time-boxed and logged.
NDA and data-processing agreement signed before we touch anything.
Every grant revoked at handoff, confirmed in writing.

QUESTIONS

Will it make numbers up?

That is the risk with any AI analyst, so it is built not to: it writes a query you can see, uses the metric definitions your finance team agrees, and every figure links to the rows it came from. If it cannot answer from your data, it says so.

How is this different from ChatGPT Enterprise or Copilot?

Those run in someone else's cloud, on a general model. This runs on your hardware or a dedicated EU server, on a model adapted to your data, and it can query your systems directly.

Ask your own question

Answered by AI from what is on this site. It can be wrong, so for anything binding, ask a human. Please don't enter personal data. Privacy

THREE WAYS IN · ONE TEAM

INFERENCEMake serving cheap.Lower cost per token, lower latency, same answers. GPU PERFMake the hardware fast.Profiling, kernels, scaling. More work per GPU-hour. PRIVATE AIMake it yours.Your own model, on your data, on your hardware.

Tell us which questions your team asks every week. We will tell you if this answers them.

GET IN TOUCH →

Local BI.Ask yournumbers.