Product · AI
Attach a language model to your database — encrypted, in place
Most ways of putting AI on your data involve sending that data somewhere else. AthenaDB is being built to do the opposite: bring the model inside the database, load it encrypted, run it in protected memory, and let your records stay exactly where they are.
The model loads encrypted
Model files are stored encrypted and signed, with their keys custodied the same way your database keys are. A package that has been tampered with does not load. Nothing is decrypted to disk.
It stays in protected memory
Weights are decrypted straight into protected memory that is locked out of swap and wiped when unloaded — and the same protection covers the working state of a running query: the prompt, the retrieved context, and everything the model produces along the way.
Your data never leaves
Because the model runs inside the database, answering a question does not mean shipping your documents to somebody else's servers. There is no outbound call in the built-in path — the retrieved context goes from your storage to the model and back without crossing a boundary.
Built in, not bolted on
The runtime is ours. That is what makes the guarantees above possible: encryption, memory protection, permissions, and audit are properties of the engine, not of a separate inference server we happen to start alongside it.
Run it where it suits you
The AI tier is often the part that wants different hardware. You should not have to give up the security model to move it.
In the database
The simplest deployment: the runtime lives in the database process. Nothing leaves the machine.
On a model machine you own
Put the AI tier on its own hardware — a GPU box, say — as an AthenaDB model node that serves inference and hosts no databases. Your database reaches it over a mutually authenticated, encrypted link between instances you control, with both sides recording what was asked.
Or an outside provider
Prefer a frontier model for some workloads? Connect one. Because that path leaves your boundary, it is opt-in, restricted by data classification, and audited — sensitive material can be blocked or redacted before it is ever sent.
AI as part of a query
Once a model is attached, language tasks become things you can ask for directly:
- Ask a question and get an answer grounded in your own records, with citations
- Classify, summarize, and extract structured fields as part of a query
- Enrich documents automatically as they are ingested
- Every call permission-checked, audited, and metered
These are kept deliberately distinct from ordinary calculations in a query: they can be slow, they cost money, they are not perfectly repeatable, and they need permission. So they are their own kind of function, with their own entitlement and their own record.