VTX Project: Autonomous 5-Layer Cognitive Architecture over Llama-3.1-8B

Hello everyone! I want to share the results of my latest project — VTX. This is a fully autonomous system deployed locally on a Linux environment (Acer Nitro V15).

While the core “engine” is Meta-Llama-3.1-8B-Instruct (GGUF, Q4_K_M) , my primary focus was building a sophisticated software orchestration layer. Instead of direct interaction with the LLM, I implemented 5 Cognitive Layers that act as a strategic controller for the model.

Key Architectural Features:

• Layered Cognitive Logic : Each layer handles a specific task — from context filtering and system prompt protection to preventing recursive “infinite loops”.

• Performance on Linux : Running on a Nitro V15, the inference is stable and fast. I’ve implemented a custom caching system that allows for near-instant context restoration in complex dialogue branches.

• Zero-External-API : The project is entirely air-gapped and independent of the internet. This is a critical requirement for my work with sensitive data, such as medical and legal information.

• Custom Visualization : I built a dedicated web interface called “Resonance Journal” to visualize the neural network’s logic and system logs in real-time.

Technical Stack:

• Model : Llama-3.1-8B-Instruct (Q4_K_M)

• Platform : x86_64 Linux (Acer Nitro V15)

• Orchestration : Asynchronous Python-based engine

• Safety : “Asymmetric caution” approach to ensure strict ethical invariants and prevent system leaks.

I am very interested in discussing multi-layered LLM management with the community. Has anyone else experimented with rigid logical filtering at the “cognitive middleware” level rather than relying solely on the model’s instructions?

Looking forward to your thoughts and feedback!