An air-gapped, zero-I/O client-driven orchestrator kernel optimized for localized unified memory deployment. Bypasses persistent cloud overhead through real-time runtime semantic compilation.
- Client Layer: Modular interfaces inject independent formatting rules directly via
ExecutionRequestenvelopes. - Kernel Routing Engine: Asynchronous Python FastAPI loop intercepting raw stream buffers. Locates routing instructions via regex look-ahead flags.
- MoE Hardware Driver: Triggers cold model parameters swaps via Ollama context clearing loops, preserving VRAM limits.
- Episodic Consolidation: Background text summary aggregation writing out asynchronously to a SQLite WAL database engine.
- **Slight context drift when hotswapping front ends if the front ends have proprietary formatting requests
- **Higher weight models tend to not use the web scraper tool provided, currently working on prompt engineering fix
##Artemis Mate Front End
- **Godot based Vtuber front end. Utilizes injection of formatting in context to allow for variable passes that control the 3D model puppet via semantic keywords.
- **Utilizes .vrm model and basic godot animation calls. As animations are canned and model is not a proprietary file type, this front end is fully moddable.
- **Base model made using basic options in VRoidStudio, with idle pose taken from mixamo.
- **Encourage looking at vgen or booth for purchase of custom assets from real artists.
**Current testing is on an M1 Mac Studio Max with 64GB of unified memory. Currently have had no hardware strain, even under inference load.
- **Create a scheduler for asynchronous tasks, allowing for active daemons to push alerts without user prompt
- **Creation of Peer to Peer (P2P) protocol for hashing user requests via a bit torrent-like protocol via SLM integration
**This system was designed for headless operation.
- **Set up your ollama instance with your desired models
- **Insert your model names and point the main bow core code at your ollama instance
- **Run the bow python as a uvicorn server
- **Point your desired front end at the uvicorn IP address you've created
- **Enjoy your low hardware spec LLM experience :)
**The system I'm running on currently is robust, so my models listed in code are heavy. I would HIGHLY SUGGEST running a swarm of 8B Q4 models with heavy specialization instead of the models currently listed in my code, unless you're masochist that likes 1 token per second. I have tested on an Nvidia 3070 and an Nvidia 1080. 8B Q4 models are the standard and work surprisingly well on 8GB of VRAM, but any higher will choke your hardware.