Overview
LVIS is an Electron-based local AI assistant platform. Users install plugins that add specific capabilities — meeting summarization, local file search and organization, document Q&A — without sending sensitive data to external APIs.
Problem
Enterprise users need AI assistance for daily workflows, but data sensitivity often rules out cloud LLM APIs. Alternatives — running models locally — require technical setup that most employees can't manage.
LVIS solves both: a desktop app that runs local models (via Ollama/vLLM) and exposes them through user-friendly, installable plugins. No cloud dependency, no technical setup for end users.
Architecture
Core platform — Electron app managing the local model runtime, plugin registry, and shared UI components. Handles model loading, context management, and plugin lifecycle.
Plugin system — each plugin is a self-contained package that declares its capabilities (tools, UI components, required models). The platform installs plugins from a registry and integrates them into the assistant interface.
Meeting summarization plugin — connects to the system audio (or meeting app integration), transcribes, and generates structured summaries (decisions, action items, attendees).
Local file search plugin — indexes the user's file system with semantic embeddings. Natural language queries return ranked file matches with context snippets.
Design Principles
Local-first — model inference runs entirely on the device. No data leaves the machine unless a plugin explicitly enables a cloud fallback.
Plugin isolation — plugins run in sandboxed contexts and communicate with the platform through a typed API. A buggy plugin can't crash the core assistant.
Incremental capability — users start with the base assistant and add plugins as needed, rather than receiving a monolithic feature set.
Status
In active development. The core platform and meeting summarization plugin are functional. Planning for deployment to internal teams for feedback before broader rollout.