01 — Workspace
Agents that stay on your compute.
Soriku is in-house AI infrastructure. Prompts and tools run on EU GPU. No third-party API sitting in the request path by default.
EU GPUPrivate inference
Walkthrough
Scroll. The screen stays put, the story moves.
01 — Workspace
Soriku is in-house AI infrastructure. Prompts and tools run on EU GPU. No third-party API sitting in the request path by default.
02 — Agents
Define agents with clear tools and memory boundaries. Inspect what ran. Change a tool without rewriting the product that calls it.
03 — Alpha
We use it ourselves before we sell it. Same rule as the rest of the family.
AI infrastructure that does not leave the building.
Spec sheet
Run inference where your data already lives. No silent hop to a US endpoint.
Register tools, scope them, keep transcripts. Agents stay inspectable.
Built to sit behind Wemazu and Simezu apps, not as a chat toy on the side.
We dogfood it. Public surface comes after the boring parts are stable.
Under the hood
Python services, EU GPU, private inference. The point is control, not a model leaderboard.
Releases
Benchmark cells are stamped with their provenance when written, and readers enrich older cells that predate the stamp. The router carries provenance through its ranked models and routing decisions, and the benchmark, model matrix and status endpoints expose it. A stale measurement now reads as stale instead of silently becoming a category of its own.
Feature updateFirst release that ships the LAN network add-on, built for macOS and Windows in the same run. The network board sits as a tab inside Connect and reads from the live engine rather than a mock. The add-on refuses hosted mode outright, so the tab simply does not exist on soriku.ai.
Feature updateThe Content Security Policy is now actually enforced instead of merely declared. A dead super administrator route was removed, and the keystore carries a test that proves it cannot leak a key. A full translation pass shipped with it.
Security updateRoles and areas for teams, with offline network invites for machines that are not on the internet. The hub turns people from HubSpot and documents from Notion into work surfaces instead of raw text. Includes a fix for the role bug in Simezu invites.
Major releaseServer, health endpoint and the About screen read one version string from the core, so the desktop app and the API can no longer disagree about what current means. The macOS download is built from the release pipeline, with parity and freshness tests.
Feature updateOur own AI infrastructure. EU compute, no third-party API in the request path.
AlphaSoriku.ai 1.0.0, rebuilt with the installer fixes in place and a disk-image step that no longer trips over Full Disk Access. A release script checks what users actually receive rather than what sits on the origin.
LaunchReal cross-platform correctness for the desktop build, plus a Windows and Linux release scaffold that is marked as untested rather than implied to work. The hosted front end is served from the release bundle.
Feature updateFour findings from an external security review are fixed, along with several holes the report did not catch. Bring your own key gained an error taxonomy, a recoverable message when a key is rejected, and a test that proves a failed key cannot leak cost. Managed API credentials carry limits and scopes.
Security updateAlso in the family
Lightweight deployment platform between Vercel and Kubernetes. Push code, pick an EU region, run.
Auth, payments and tenant management. JWT with refresh rotation, OAuth2/OIDC, Mollie.
