Skip to content

Introduction

DUUMBI is an AI-first semantic graph compiler. Instead of storing programs as text files, DUUMBI represents program logic as typed JSON-LD semantic graphs and compiles them directly to native binaries via the Cranelift backend.

Programs should be structured, verifiable data - not strings of characters that require parsing.

In DUUMBI, a function is not a sequence of characters. It is a graph node with typed edges to its parameters, body blocks, and operations. This means:

  • No syntax errors - the graph structure is valid by construction
  • AI generates programs directly - LLMs produce structured graph patches
  • Every mutation is validated - schema validation and type checking happen at the graph level
JSON-LD source -> semantic graph -> validation -> Cranelift IR -> native binary
  1. JSON-LD source - Programs are stored as .jsonld files using the duumbi: namespace
  2. Semantic graph - Parsed into a petgraph::StableGraph, the single source of truth
  3. Validation - Schema validation, type checking, and ownership verification
  4. Cranelift compilation - Graph nodes lower to Cranelift IR, one function per subgraph
  5. Native binary - Object file links with the DUUMBI runtime
  • Semantic Graph IR - petgraph-backed StableGraph as the program representation
  • Query-first workflow - ask about the workspace without mutating it
  • AI-native mutations - explicit Agent and Intent handoffs for write-capable work
  • Provider catalog - direct provider keys for Anthropic, OpenAI, xAI, MiniMax, DeepSeek, Qwen, Moonshot, Zhipu, and Gemini
  • Cranelift backend - native compilation without LLVM
  • Intent-driven development - describe what you want and let DUUMBI plan it
  • Ownership model - Rust-inspired borrow checking at the graph level
  • Module registry - publish and install graph modules with SemVer resolution

DUUMBI separates read-only exploration from write-capable mutation:

Ask a question
-> Query mode
-> answer with sources
-> optional explicit handoff to Agent or Intent

Query mode may inspect, explain, compare, and recommend. It does not apply graph patches or write files. Agent and Intent mode are explicit handoffs for mutation or planned implementation.