Heidi Unified Learning Suite

A self-improving, modular AI suite for local model hosting, runtime learning, and offline dataset curation.

Note

Heidi-CLI has been evolved into the Unified Learning Suite. Legacy agent orchestration features (Copilot SDK, Jules, OpenWebUI) have been removed to focus on a local-first learning loop.

🚀 The New Learning Loop

Heidi now provides a complete feedback loop for local AI development:

1. Multi-Model Host: Serve multiple local LLMs via an OpenAI-compatible API.
2. Runtime Learning: Models gain experience through memory, reflection, and reward scoring.
3. Data Pipeline: Raw runs are captured, redacted, and curated into training datasets.
4. Self-Improvement: Background retraining, evaluation gates, and atomic hot-swapping.

🛠 Modules

• model_host/: OpenAI-compatible multi-model routing.
• runtime/: Memory layers, reflection, and strategy ranking.
• pipeline/: Data capture, curation, and secret redaction.
• registry/: Versioning, eval gates, and promotion logic.

📦 Getting Started

Installation

python -m pip install -e .

Initialize the Suite

heidi status

This will initialize your persistent state root (default: ./state).

Serve Models

heidi model serve

The host will listen on http://127.0.0.1:8000 and provide:

• GET /v1/models
• POST /v1/chat/completions

🩺 Doctor

Verify your installation and modular structure:

export PYTHONPATH=$PYTHONPATH:$(pwd)/src
python3 scripts/doctor.py

📖 Documentation

• Architecture
• Model Host
• Auto-Registration

---

Heidi is building the future of local, autonomous AI learning.