TrinityCoordinator
An Elixir and Nx implementation of a TRINITY-style small-model router that reads a conversation, extracts a hidden-state vector, and routes each turn to the right agent and role before dispatching it to a live LLM provider.
trinity_coordinator is an Elixir and Nx runtime for a TRINITY-style local router. A compact Qwen3-0.6B model, adapted through singular-value fine-tuning, reads a transcript; the runtime extracts a hidden-state vector, a small router head maps it to agent and role logits, and a configurable provider boundary dispatches the selected work to a live LLM provider.
The repository contains the Elixir runtime, parity tooling used to validate it against the original Python pipeline, and Mix tasks for onboarding a fresh contributor from a clean clone. The runtime path from router to role injection to provider boundary to traced turns works end-to-end with deterministic mock providers and gated live-provider calls, and its evaluation suite passes 37 of 37 tests on CUDA.
The project targets CUDA, Apple Silicon and CPU runtime profiles, is still in active development, and a clean clone resolves its sibling dependencies from GitHub with no other local repositories needed.