Patois.

The IT brief

Localization infrastructure your security team can approve.

Localization OS runs entirely inside your environment: your servers, your storage, your models, your rules. Here is how it is put together, and exactly where every control stands today.

Architecture

One process. Your data path, end to end.

Clients talk to a single application. The application talks to your database, your file storage, and the model provider you configure. Nothing routes through Patois Labs.

Localization OS deployment architecture Clients (web UI, CLI, and API or MCP agents) connect to Localization OS, one process built as a modular monolith. Inside it: translation memory and glossaries, LLMemory retrieval, workflows with LQA, the CAT editor, the LLM Ledger, and connectors, on top of named ports for identity, models, storage, scheduling, and telemetry. The application talks to your data store (SQLite by default, PostgreSQL supported), your file storage, and your identity provider over SAML 2.0 and SCIM 2.0, proven against Okta. Model traffic goes only where you configure it: fully offline local models, your internal endpoints, approved frontier APIs, or a hybrid chosen per workflow. No Patois service is in the data path. Web UI your team, in a browser CLI headless and CI API · MCP agents everything, scriptable Localization OS one process · a modular monolith TM · glossaries LLMemory (RAG) Workflows · LQA CAT editor LLM Ledger Connectors NAMED PORTS Identity · Models · Storage · Scheduler · Telemetry Your data store SQLite by default PostgreSQL supported Your file storage uploads · assets · keyfiles Your identity provider SAML 2.0 · SCIM 2.0 proven against Okta Model traffic goes only where you configure it A · Fully offline local GPU · LM Studio-style B · Internal endpoints your own model servers C · Approved frontier APIs your keys, your terms D · Hybrid chosen per workflow

Model traffic follows the customer's configuration. There is no Patois-side service in the data path.

Controls

The controls.

Deployment

Single node. No orchestration required.

Runtime
One FastAPI process, shipped as a Docker image or run as plain Python. There is no Kubernetes to stand up and no service mesh to babysit.
Topology
One node is the whole story: a laptop, a VM, or an air-gapped box that has never seen the internet. Patois Labs runs its own production install exactly this way: one container on a single AWS instance, behind Cloudflare and a TLS proxy.
Data store
SQLite by default, which means zero database operations on day one. PostgreSQL is supported when your platform team wants it.
Hardware
Modest. A GPU matters only if you host models locally; pointed at an API, it runs on almost anything.

Identity and access

SSO and provisioning, proven against a real IdP.

SSO
SAML 2.0 single sign-on, interop-tested against a real Okta organization. A live round trip, not a checkbox on a datasheet.
Provisioning
SCIM 2.0 user lifecycle: users and groups created, updated, and deprovisioned, exercised against that same Okta org.
Authorization
Role-based access control separates who can read, translate, review, and administer.
Sign-in
Two-step sign-in with an authenticator app and recovery codes, which you can require for everyone; a password policy; account lockout after repeated failures; and an absolute session limit as well as an idle timeout.
Headless
Per-user API keys for automation, CI, the CLI, and MCP agents, governed by the same permission model as the people.
Rollout
Enterprise features ship flag-gated and off by default. Nothing turns on until you turn it on.

Audit and data protection

Every action recorded, everything at rest encrypted.

Audit
Every action lands in an append-only, tamper-evident audit chain, and the chain itself is monitored. If history gets touched, it shows. Deleting the newest rows is caught too, not just edits.
Key rotation
The audit signing key lives outside the database and can be rotated on the host. History signed by the old key still verifies.
Encryption
Data and key material are encrypted at rest.
Transport
Hardened response headers and cookie flags, with HSTS verified on the live deployment.
Data rights
GDPR-grade erasure and export are built into the data model, not bolted on afterward. Erasing a person anonymizes them everywhere work is attributed. The one stated exception: audit entries written before the erasure keep what they recorded, because rewriting them would break the chain. Only administrators can see them.
Recovery
Full-install backup with a documented restore path. Disaster recovery you can rehearse, not just cite.

Supply chain and updates

You review and deploy on your own schedule.

Scanning
Every change passes a pre-merge gate that scans every locked dependency for known vulnerabilities (pip-audit), checks the native libraries bundled inside packages against an advisory list, and applies static security rules. A new advisory blocks the merge until it is fixed or recorded as an accepted exception with a reason.
Images
Release images for installs Patois Labs hosts are scanned before they serve traffic.
Delivery
Updates arrive as source through your private repository channel. You can read every diff before it goes anywhere near production.
Control
You review, you schedule, you deploy. Nothing in your environment ever auto-updates.

AI controls

The model works under your rules.

Refusals
When a model refuses or answers with chatter instead of a translation, the output is quarantined for review. It is never stored as a translation or written to translation memory.
Prompts
Workflow prompts are visible and editable, locked until someone unlocks them, applied deliberately, and restorable to the default. Every change is audited.
Transcripts
Agent runs keep a transcript of the plan and every tool call. Customer text and model reasoning in it are stored only as fingerprints (a hash and a length), never as readable text.
Spend
A monthly spend cap and per-department budgets refuse new paid calls once reached. Local models on your own network cost nothing and never trip the cap.

Connectors and agents

Outside systems and AI agents, on a short leash.

Write-back
Connectors write each language as the source system's own language version, as a draft where the connector supports drafts, and never overwrite the source. Text that failed review is held back, never written as a translation.
Egress
Connector hosts are checked before every call. Loopback and internal addresses are refused unless an administrator allowlists a specific private address. A connector also refuses to write placeholder output from the offline test engine to a real system.
Credentials
Connector credentials are encrypted at rest and never shown back once saved.
Agents
MCP tools are filtered to what the calling key may do. A paid call needs a plan and a confirm bound to that plan, redeemed once. Text from outside systems is marked as data, never instructions, and vendor errors are summarized with anything that looks like a credential removed.
MCP sign-in
An optional OAuth 2.1 sign-in for MCP clients (PKCE only, tokens bound to this install), off by default.
Kill switch
An administrator can switch off any connector kind on the install. It disappears from every surface until it is switched back on, with its connections kept.

Accessibility

Built to WCAG 2.2 AA.

Standard
The product is built to WCAG 2.2 Level AA, in both the light and dark themes.
Checks
Automated browser checks run the axe rules on the main pages in both themes, test reflow and keyboard focus, and fail if anything gets worse.
Evidence
An Accessibility Conformance Report (VPAT 2.5) and a completed HECVAT are available on request, with known exceptions listed honestly.

Telemetry

None.

The software phones home to no one.

No analytics, no usage beacons, and no license callbacks reach Patois Labs. What runs in your environment stays in your environment.

Compliance posture

SOC 2 and ISO 27001: not yet certified.

Formal gap assessments are complete against both standards, the controls are documented, and the remediation roadmap is active. Ask, and I will walk you through exactly where we stand.

SOC 2 · gap assessment complete ISO 27001 · gap assessment complete Remediation roadmap · active

Continuity

If Patois Labs disappears, you keep running.

Delivery is the source itself, which makes escrow built in. You hold the code, your license, and a system already running inside your environment. There is no vendor lock to unwind and no server to lose access to. The software keeps working whether or not Patois Labs does.

Bring your architect.

Book a technical walkthrough and put the whole system in front of the person whose signature you need.

Schedule a technical walkthrough →

or email brent@patoislabs.com