Security & Architecture
Alethiom deploys in your cloud account. We operate it through a role you grant — and can revoke.
That’s the pattern under review. This page enumerates the rest: what runs in your account, what moves and what never moves, and the controls that stay in your hands — the logs you can read, the access you can revoke.
Data flow
Where data lives, what moves, what never moves.
There is no vendor-side data store: customer data is never copied to an Alethiom cloud. The whole product runs inside your account — the UI your team opens, the APIs behind it, the AI service, the sync workers — so query results flow from your cloud straight to your users’ browsers. One outbound channel reaches Alethiom — configuration, health, and billing metering, never customer data.
data plane
Event sourcesvendor exports · CDP · server-side streams — however your pipeline works today
your cloud account
Your warehousesystem of record
Alethiom deploymentengine · UI · APIs
sync workers
Your team’s browsersquery results, passed securely
what moves: bounded, scheduled sync queries from warehouse to serving layer, using your warehouse’s native CDC capabilities (inside the account)
what never moves: customer event data, to Alethiom or anyone else
Deployment model
One stack, one role, revocable at any time.
- You run one infrastructure stack from us. It creates one scoped role in your account, granted to us — that role is the entire surface of Alethiom’s access.
- Through that role, we build and operate the Alethiom stack inside your account: provisioning, upgrades, monitoring, incident response. You never operate our software; you host it.
- You can revoke access at any time. Revocation cuts our access immediately; your data was never anywhere else to begin with.
- Ships as
- Alethiom ships one way: managed, inside your cloud account — a standard Kubernetes cluster running the full stack, visible to your own audit and security tooling. AWS and GCP today; Azure coming soon.
- Warehouse access
- Snowflake and BigQuery today; Databricks coming soon. A read-only role you grant, scoped to the datasets you choose. Only scheduled sync queries touch the warehouse — native CDC, as often as every 15 minutes, with record-level appends, edits, and deletes all supported. Interactive queries never hit your warehouse.
- Operations
- Our access runs through roles and credentials that you grant and can revoke; your cloud’s audit logs record our infrastructure actions.
- Encryption
- TLS in transit; at rest under your cloud’s storage encryption. The warehouse credential is stored encrypted inside your account and never leaves it.
The stack
What runs in your account.
Everything below runs inside a Kubernetes cluster inside your account:
- ClickHouse
- The serving layer: open-source, columnar — a disposable cache built from your warehouse, stored in sequence order and sharded by person, with sessions, stitching, and attribution computed at query time. Rebuildable at any time; the warehouse stays the system of record. Deletes and updates propagate on the next sync — deletion and update requests are honored in one place: your warehouse.
- Postgres
- Metadata: definitions, boards, saved work, configuration.
- Keycloak
- Authentication, federating your identity provider — your SSO, your MFA policy, your deprovisioning. Alethiom holds no separate password store.
- Application services
- The UI your team opens, the APIs behind it, the AI service, the sync workers.
AI posture
AI on your terms — including “off.”
The AI teammate is opt-in with a real kill switch. Not buried, not defaulted on, not “contact support to disable.”
AI off
Zero model-provider traffic; the only outbound channel is the operations endpoint.
Everything else in the product works exactly the same.
AI on
Model calls run under commercial API terms that are processing-never-training — your data is never used to train anyone’s models.
Prefer your own agreement? Bring your own Anthropic key: AI runs under the model-provider contract you negotiated.
When AI is on, every answer it gives is a traceable query — the same governed layer your team uses, with the same permissions as the person asking.