Project Apollo: Sovereign AI OS

An air-gapped, local MoE swarm and routing architecture designed for absolute human agency, bypassing centralized data aggregators and cloud-compute monopolies.

---

🛑 The Mission

Project Apollo is a "Sovereign AI" operating system built on the premise that if you don't own the hardware, you don't own the truth. It is designed to run 30B+ Mixture of Expert (MoE) models, Vision models, and RAG pipelines entirely locally on consumer-grade hardware.

The goal is absolute privacy, deterministic execution, and immunity from the "Censorship Tax" and "Cloud Tax" imposed by proprietary API providers.

🏗️ System Architecture ("Three-Mind / VRAM Tetris")

Apollo uses a cascaded, intent-based routing system to dynamically load and unload specialized models into memory based on the task, maximizing the utility of a strict 16GB VRAM ceiling.

1. Gatekeeper (System 1): qwen3:0.6b (Resident). Handles fast triage, intent classification, and chitchat.
2. Engineer (System 2): qwen3:8b (On-Demand). The primary logic workhorse for 90% of local tool execution and context gathering.
3. Architect (System 1.5): qwen3-coder:30b (On-Demand). Complex structural logic, CAD design, and deep reasoning.
4. Reasoning Specialist: deepseek-r1:14b (On-Demand). High-fidelity chain-of-thought logic.
5. Vision: qwen3-vl:8b (Native ROCm). Multi-modal physical world and desktop analysis.

📂 Core Features & Achievement Proofs

1. Multi-Tier Agent Orchestration

Architected a three-stage cascading dispatcher (modules/router.py) that programmatically triggers Python modules based on natural language intent, optimizing token generation speed vs. model parameter depth.

2. VRAM Resource Management ("VRAM Tetris")

Engineered a strict hardware-aware protocol (vram_management.py) that monitors GPU health and verifies memory release before allowing the router to trigger a model swap, eliminating kernel panics on the ROCm 7.2 stack.

3. The "Librarian" Local RAG

An autonomous data ingestion pipeline (librarian_ingest.py) that scrapes URLs and PDFs into a local ChromaDB Vector Database for semantic retrieval, providing a completely private knowledge base.

4. The Forge

A structured two-layered system (modules/forge.py) for capturing raw engineering visions and autonomously refining them into executable project proposals using the 30B Architect.

5. Discord UI Bridge

A custom Discord bot integration (discord_bridge.py) acting as the primary UI, featuring real-time hardware telemetry dashboards, image ingestion for physical inventory audits, and an interactive security approval queue.

🛠️ Hardware Requirements

• Primary Host: AMD Radeon RX 9070 XT (16GB VRAM) / Ryzen 7 5700X3D
• Stack: Ubuntu 22.04 / ROCm 7.2 / PyTorch 2.12.0.dev
• Backend: Ollama (Native ROCm) / vLLM

🚀 Getting Started

1. Clone the repository.
2. Set up your .env file based on .env.example.
3. Ensure ROCm 7.2 is configured on your local machine.
4. Run python3 apollo.py or start the Discord bridge with python3 discord_bridge.py.

---

Developed by Mark | Lead AI Systems Architect