Research
Notes on hybrid code : Modular AI, micro-transformers as plugins, mixed C++ and neural inference behind a shared plugin host. Separate from the production MCP stack.
The production components these notes refer to, the model backend, the shared VRAM pool and the RAG pipeline, are documented under AI Hub. Hardware-level work, firmware analysis and DSP programming, is collected under Reverse Engineering.
VRAM Allocation for Modular Neural Plugins
How the Neural-layer VRAMManager hands out GPU buffers to ITensorModule plugins : RAII handles, prioritised LRU eviction, and refcounted shared weights.
Modular AI : Micro-Transformers as Hybrid Plugins
Treating small specialised models as plugins, mixed with C++ code, behind a shared plugin host : Shared VRAM, embedding bus, KV cache pool, training hooks.