We’ve been told a lie.
The corporate gatekeepers of the artificial intelligence boom want you to believe that intelligence requires a centralized mainframe, a billion-dollar data center, and a permanent connection to their cloud. They want you to believe that in order to query your documents, you must first surrender them to their servers—feeding the Matrix with your patients' private clinical histories, your clients' legal secrets, and your own personal thoughts.
We reject this.
ANDARTIS is an experiment in sovereign intelligence. It is a local-first, air-gapped framework built on a hybrid architecture of a NativePHP desktop app and an Apple MLX acceleration engine. It is designed to run on standard, consumer-grade Apple Silicon—a 16GB RAM democracy where hyper-specialized micro-brains and quantized SLMs communicate locally via a zero-latency JSON-RPC STDIO pipeline. No web servers. No external API keys. Zero data egress.
Today, the ANDARTIS experiment is officially out on GitHub. Let’s talk about how it works, why it is built this way, and who it is for.
The Core Design: "How to Think" vs. "What is Fact"
To make local-first AI viable on a standard laptop, we had to solve the twin challenges of VRAM limitations and data privacy. We did this by strictly separating model adaptation from context ingestion.
+---------------------------------------+
| 🧠 1. MODEL FORGE: "How to Think" |
| Anonymized datasets -> MLX LoRA |
+---------------------------------------+
|
v
+-------------------------------+ +-------------------------------+
| 💾 2. ACTIVE NODE: "What is | | ⚡ LOCAL INFERENCE ENGINE |
| Fact" | ------> | Synthesizes answers locally |
| Raw Files -> SQLite Core DB | | without data egress |
+-------------------------------+ +-------------------------------+
|
v
+-------------------------------+
| 🖥️ NATIVEPHP DESKTOP UI |
+-------------------------------+
1. The Model Forge: How to Think (LoRA Adaptation)
The local Small Language Model (a quantized Mistral-7B) needs to understand your domain. A doctor uses clinical records; a lawyer uses discovery files; a business analyst uses corporate invoices.
We do not train the model directly on active, sensitive files. Doing so risks overfit-memorization, which is both computationally heavy and a potential local privacy hazard if model weights are ever shared. Instead, we use the Model Forge. You feed the forge anonymized, generic style templates (e.g., sample blank medical charts, standard legal brief formats).
Using Apple's MLX framework, we run a local LoRA (Low-Rank Adaptation) fine-tuning run. The model learns the grammar, the jargon, and the structure of your domain. It learns how to extract entities and compile layout blueprints.
2. The Active Node: What is Fact (SQLite Ingestion)
The actual operational files—active patient charts, active litigation papers, confidential journals—are never fed into the neural network's weights.
Instead, they are ingested into an Active Node. The ANDARTIS Librarian walks your chosen local folder, chunks the text, and stores the raw structured facts inside an isolated, dedicated core.sqlite database file on your disk. When you query a node, the local inference engine fetches the exact facts from SQLite and feeds them as a grounded context block to the resident SLM. The model answers using only the provided facts, ensuring zero hallucinations and absolute data sovereignty.
The Engineering Breakthrough: Millisecond "Hot Syncs"
Fine-tuning a LoRA model and calculating vector embeddings for thousands of text chunks is a resource-intensive task. If the application had to re-process and re-train every capability from scratch every time you edited a file, it would be unusable.
In our latest repository updates, we implemented an elegant caching and change-detection pipeline:
- File Idempotency: The ingestor now hashes files and checks the isolated SQLite core. Unchanged documents are skipped in less than 1ms.
- Vector Index Caching: The
SemanticNavigator compares the database's current chunk IDs against the vector index metadata. If they match, it skips generating embeddings.
- LoRA Caching: The
PatternSynthesizer checks if the safetensors adapter is already aligned with the active chunks. If so, it bypasses the mlx_lm lora training process entirely.
The result is a system that cold-syncs a new folder in minutes, but hot-syncs subsequent updates in milliseconds.
Who is this for? The Sovereign Thinkers
ANDARTIS is not built for corporations seeking to optimize their slide decks. It is built for the rebels, the independent practitioners, and the sovereign thinkers who operate outside the lines of massive institutional systems.
1. The Rebel Doctor
Imagine a cardiologist running private clinical research, searching for correlations between obscure medication side effects and patient outcomes. Institutional review boards and corporate hospital networks make it impossible to run these searches using public cloud tools because of compliance lockouts. With ANDARTIS, the doctor drops patient summaries into a local folder, boots a cardiology LoRA adapter, and queries the clinical nodes locally on an offline MacBook. Complete data security. Zero administrative friction.
2. The Independent Psychologist
A therapist wants to map cognitive loops, emotional triggers, or recurring themes in thousands of pages of patient diary logs. Uploading these deeply intimate human stories to OpenAI is an ethical violation. By running ANDARTIS, the psychologist processes the files locally. The system builds semantic relationship matrices and extracts key themes entirely inside their home office, respecting the sacred trust between therapist and patient.
3. The Sovereign Lawyer
An independent attorney taking on a massive corporate entity is buried under 50,000 pages of discovery documents. They need to find obscure regulatory infractions, antitrust violations, or corporate negligence buried in email threads. Large corporate firms throw armies of paralegals at the problem. The independent lawyer runs a federated cross-reasoning query across their local nodes. The Intent Blade translates their natural language search into exact SQLite queries, matching legal infractions with 100% mathematical precision.
Getting Off the Grid
The current AI landscape wants us to believe we are helpless without their clouds. But the hardware in your backpack—that M-series Mac—is incredibly powerful when combined with low-level, metal-optimized code like Apple MLX and NativePHP.
ANDARTIS is an open, sovereign experiment in taking back control of our own thoughts.
The code is live. Bring your own datasets. Forge your own brains.
👉 Join the Experiment on GitHub