Back to projects
Atrium

Atrium

In Development

Creator, Architect, AI Solutions

Atrium is a desktop application that runs specialized AI assistants directly on the user's machine, with no cloud. Unified UI in Tauri (Rust + Svelte 5 / TypeScript), Python backend driven over WebSocket, local inference via Ollama (qwen3 models). Each persona has an agentic anti-hallucination RAG: tool-driven search over a SQLite + FTS5 index with bge-m3 embeddings, citations limited to what the tools actually return, deterministic short-circuits, per-persona fact memory. Local voice synthesis (Kokoro). Built local-first and offline-capable, with a selector to switch between assistants.

What this project proves

7Sail R&D · Local-first Tauri app · Anti-hallucination RAG

Current state

Private development • internal Tauri build • no public release

Design focus

TauriRustSvelte 5

Context

Internal 7Sail production bet: a sovereign AI workstation that runs offline on the user's machine, where several specialized assistants coexist behind a single interface. The core stake: reliability, assistants that do not hallucinate and only cite what they actually read.

Challenges & constraints

Guaranteeing anti-hallucination on constrained local models (citations limited to tool output, automatic correction, deterministic short-circuits). Letting multiple personas coexist under a single UI with a shared LLM runtime. Holding a memory / VRAM budget on local hardware while keeping response times acceptable.

Results & learnings

Functional desktop app (installable Tauri build), shared Ollama runtime, tested WebSocket contract, agentic RAG operational on a first corpus, local voice synthesis. Reusable persona architecture (shared contract, mutualized bge-m3 RAG) to plug in new assistants without rebuilding the foundation.

Tech Stack

TauriRustSvelte 5TypeScriptPythonOllamaqwen3bge-m3SQLite FTS5WebSocketKokoro TTS

Features

  • Specialized AI assistants, 100% local (Ollama)
  • Agentic anti-hallucination RAG (SQLite + FTS5 + bge-m3)
  • Citations limited to actual tool output
  • Per-persona fact memory
  • Local voice synthesis (Kokoro)
  • Unified multi-persona UI
  • Local-first / offline-capable
  • Tauri desktop app (Rust + Svelte)

Production note

The heart of the project is anti-hallucination: on local models, an assistant only has value if it cites only what it actually read. The whole foundation (tool-driven RAG, short-circuits, automatic citation correction) is built around that constraint before anything else. The unified UI and shared runtime then capitalize on it to host multiple assistants without duplicating infrastructure.