Give your business
one brain.
Before you choose tools, models or agents, build the layer they all need. Everything plugs in, swaps out and grows as your business does.
Book 30 minutesEveryone is racing to add AI agents. Almost no one is building the foundation underneath them.
An agent is only as good as the context it can see. Point one at your business today and it acts on fragments: one system, one dataset, no memory of why anything was decided. Pragma is the layer that gives every tool, agent and model the full context of your business.
Senior time, spent answering
Your most expensive people, interrupted for answers that are already written down.
AI is already in the building
Your people already paste company information into personal AI accounts. Unshared, unmanaged, invisible to you.
Knowledge that walks out
How things really work lives in a few heads. When one leaves, it leaves too.
The gap compounds quietly
Businesses that run on AI get faster every month. The distance between you and them never announces itself.
A customer repeats themselves for the third time while your agent digs through old tickets.
With Pragma: the full history and the right policy, in front of your agent in seconds.
Every proposal starts from a blank page while the answers that won last time sit in old bids nobody can find.
With Pragma: past bids, pricing and case studies surfaced in seconds, so you reuse what already won.
Your engineering lead says it will be ready in a month. The only people who can verify that are the ones on the hook to deliver it.
With Pragma: an independent, expert read on real progress, and a team that ships more each sprint because nothing has to be re-learned.
Your new hire's first month is spent interrupting your best people with the same questions.
With Pragma: they get the answer a five-year veteran would give, without asking one.
Every decision still routes through you, because you're the only one holding the full picture.
With Pragma: the business answers its own questions, and you get your week back.
It doesn't replace anyone. People still make every decision. Pragma just makes sure the right information is in front of them when they do.
Pragma is a set of containerised services that install into your environment: one ingestion and retrieval layer, four ways to consume it, and a single gateway that every model call passes through. Nothing here is a wrapper around someone else's chat product.
pgvector, the embedding pipeline, and a small local model for query translation. The part that decides what the LLM is allowed to see.
Everything inside the shaded boundary runs on infrastructure you control. Only outbound model calls cross it — and only the ones you configure.
This is where most internal AI projects stall: someone works out that the knowledge has to be gathered, cleaned and moved first, and the project becomes a data migration nobody agreed to. Pragma reads your repositories as they are. No restructuring, no re-tagging, no extraction exercise. Point it at a Confluence space or a GitHub organisation and it syncs from there, incrementally, keeping pace as things change.
Retrieval is the part that decides whether any of this works. A pipeline that looks impressive across ten documents behaves very differently across two hundred thousand, and that gap is where most deployments quietly fail. We use established practice alongside techniques of our own, and it is where the bulk of our engineering has gone.
Most enterprise AI spend goes on frontier models doing work that never needed one.
Summarising a ticket. Pulling fields out of a document. Drafting a routine reply. Answering a policy question that is already written down. That is the bulk of real enterprise volume, and it is billed at frontier rates because there is nothing sitting in the path deciding otherwise.
Pragma is that thing in the path. Every call in the business routes through one gateway, and the policy is yours to set — per team, per task type, with budgets, caps and usage accounting attached. High-volume, low-complexity work goes to a local or low-cost model. The hard tail still goes to the frontier.
marginal cost only
We will not tell you a 27B model matches a frontier one. On hard reasoning and agentic coding it does not, and anyone claiming otherwise is selling you a benchmark. The claim is narrower and more useful than that: most of your volume is not hard reasoning, and paying frontier rates for the part that is not is the largest avoidable line item in an enterprise AI budget.
You do not need to choose before a pilot. Which model fits comes down to three things: how sensitive the corpus is, whether you would rather spend up front or monthly, and whether you want to run GPUs at all. Moving between them later is a deployment change, not a migration — the stack underneath is the same in each.
No migration project
Connect repositories in place and sync. Live in weeks, not quarters.
Genuinely model-agnostic
One gateway for every call. Change provider, add a local model, or run several at once.
Runs in your infrastructure
Self-hosted core stack. Optional redaction before any call leaves your network.
Answers you can verify
Every response cites the documents behind it, linked to the system of record.