- •Building an adaptive runtime that runs quantized LLMs locally on 4GB-RAM laptops, for developers without high-end GPUs who want a real, local taste of tools like Claude Code and Codex
- •Shipped real-time RAM, CPU and battery profiling plus a per-turn safety floor that halts inference before memory exhaustion instead of letting the OS swap
- •Integrating quantized model loading (GGUF via llama.cpp, Q4_K_M, mmap-based weights) behind a pre-flight capacity check that refuses a model the machine cannot hold
- •Treating the memory budget as an explicit, checked constraint rather than something discovered by crashing - the same discipline the agent repos apply to context and state
Pythonllama.cppGGUFQ4_K_MmmapMemory budgeting