The whole point is that your data never becomes someone else's input. Here is exactly what touches what.
Container image delivered on media or via your artifact registry. No outbound connections exist in the image. Verified with a network-isolated test before hand-off.
Runs in your datacenter or office server. Optional outbound only for your own monitoring stack. Updates pulled by you, on your schedule.
Your AWS/GCP/Azure/Yandex/VK/Selectel account, your VPC, your keys. We never hold credentials past the engagement.
| Phase | Option A, standard | Option B, fully private |
|---|---|---|
| Data preparation & labeling | Our engineers may use commercial AI tooling on anonymized / synthetic / non-sensitive samples, under zero-retention terms, from an account you control | All tooling is open-weights and runs inside your environment; nothing leaves |
| Model training | In your environment or on a training server we rent in a region you approve, with data encrypted at rest and wiped after delivery, your choice, in the contract | |
| Inference (production) | Always local. No telemetry, no callbacks, no license server. | |
Option B may cost a little accuracy on some tasks and is the default for healthcare, defense-adjacent and Russian-market clients.
We are engineers, not lawyers, but this is why clients come to us.
Local processing removes the cross-border transfer question entirely. AI Act obligations for general-purpose models apply from Aug 2025; high-risk obligations are being phased in (currently slated for late 2027). A narrow, documented, locally evaluated model is far easier to put through a conformity assessment than an opaque third-party API.
With inference on-prem there is no third party in the data path, so no BAA is needed for production. Audit trail, access control and logging stay in your existing controls.
Personal data must be stored and processed in-country; fines for leaks now reach a percentage of revenue. Draft AI legislation (published 2026) targets restrictions on foreign "cross-border" AI services. A model that runs inside your contour, with an open-weights base, sidesteps both.
Cisco's 2025 privacy benchmark: 64% of organizations worry about sharing sensitive information with generative AI tools; roughly half admit employees already do. Removing the API removes the shadow-AI problem for that workflow.
Send it. Answering it is usually the fastest first step.