Deodat-Lawson/LaunchStack

What it solves

LaunchStack provides a modular, TypeScript-based engine for building AI-native applications. It solves the complexity of wiring together disparate AI services by offering a standardized, ports-based architecture for document ingestion, OCR, RAG (Retrieval-Augmented Generation), knowledge graph management, and background job processing.

How it works

The project is structured as a pnpm workspace containing a core engine and a set of "bricks" (independent packages for persistence, LLM calls, orchestration, and retrieval). Developers use createEngine(CoreConfig) to initialize the system, which then manages database connections, storage, and AI provider configurations. A reference Next.js application demonstrates how to host these components, using a worker process to handle asynchronous tasks like document ingestion via a transactional outbox.

Who it’s for

This is designed for software engineers building complex, AI-integrated web applications who want a structured, maintainable way to manage AI infrastructure without tightly coupling their business logic to specific AI providers or frameworks.

Highlights

• Modular Architecture: Separates the engine core from product-specific "pipelines" and host applications.
• Comprehensive Tooling: Includes built-in support for OCR, hybrid RAG, entity extraction, and multi-agent collaboration.
• Infrastructure Agnostic: Uses a ports-based design that supports various LLM providers (Gemini, Azure OpenAI, Ollama, etc.) via the OpenAI chat-completions protocol.
• Production-Ready Patterns: Implements robust features like transactional outboxes for background jobs, audit logs, and HMR-safe singleton slots for state management.