Prompts and data leave the building
Every call to an outside AI service takes your data somewhere you cannot see.
Precepta · The control plane
One control plane for every AI request your company makes.
Precepta sits in front of every model your organisation uses. Each request is checked against your rules, routed to the right model, answered from cache when it repeats, stripped of sensitive data and recorded. It runs on your own servers, and nothing leaves your network.
The problem
Three gaps between a pilot and production.
No consistent rules across teams
Each team writes its own limits into prompts, which the AI can be talked around.
No answer when the auditor asks
What was asked, what the rules said, what the AI saw — none of it can be pieced back together.
What it does
The rails are fixed. The routing is yours.
There is no way round it. Every call takes the same checked path, before and after it reaches a model.
Firewall → Sensitivity → Policy → Smart Route → Cache → Inference → Audit
Nothing can leave
A signed proof that nothing left, plus continuous checks that the exit really is closed.
A record that cannot be quietly changed
Every request, decision and answer is recorded and linked. That is your answer when compliance asks what happened.
Hard limits
Limits per API key, and caps on cost and usage. They sit outside the model, so it cannot talk its way around them.
Smart routing
Set the model to auto and it picks the best one available, balancing quality, cost and speed.
Data protection
Personal data is removed before the model sees it, and the answer is checked for leaks on the way back.
Change one line
It speaks the OpenAI API, so you point your existing code at it and change the address. Nothing else.
How it runs
Not a setting you trust. A property you can prove.
Not "we look after your data carefully" — there is nowhere of ours for it to sit. The models, the gateway, the rules and the records are all on your servers.
Self-hosted
Runs inside your cloud account or data centre, with no public route. Nothing reaches us — there is no telemetry and no phone-home.
On-premises
Including networks with no egress at all. The same artifact runs in both, so what you evaluate is what you operate.
Air-gapped
Networks with no outside connection at all are what this was built for, not an awkward exception.
Zero-egress is proven where it matters, not on a web page.
It needs the software inside your network, so it is demonstrated during an evaluation on
your own infrastructure — with your security team choosing the tests.
What being in the path is worth
The same position that governs a request can also make it cheaper.
Anything that sees every request can do more than check it. Precepta is already in that position, so the controls that make AI defensible also make it cost less to run.
It picks the model
Ask for auto and the smart router sends the request to the model that can
answer it inside your boundary. Not every question needs your most expensive one.
It answers twice for the price of once
Deterministic answers are cached. A request served from cache is a request you did not pay to run again.
It holds the ceiling
Prompts are compressed before they are sent, and every key carries its own scopes and spending caps. Nobody finds out about a runaway job from the invoice.
Who it's for
CIOs, CISOs and platform teams.
The people who have to answer for what the AI touched — and who already own the servers it should run on.
How we price it
By environment, throughput and retention.
Because nothing leaves your network, we cannot meter your usage remotely. Capacity is agreed up front and checked against your own logs. That comes from how it is built, not from a gap in our billing.
What you get
Running today, in your boundary.
| Capability | Deployment |
|---|---|
| Llama 3.2 in-boundary, Smart Router, local Nomic embeddings, OpenAI-compatible API | Self-hosted · on-premises · air-gapped |
| Firewall, sensitivity, policy, PII redaction, output leak scanning, cache, audit | Self-hosted · on-premises · air-gapped |
| Signed sovereignty attestation, zero-egress verification via live probes | Self-hosted · on-premises |
| No public route, no telemetry, no phone-home | Self-hosted · on-premises |
Put it in front of your security team.
We set it up on your own servers and your team picks the tests — including trying to get data out.
Every model call is now traced. Intent Studio governs what work they were doing in the first place.