Uncover the low-level architecture that powers our sovereign runtime. From native JIT KV-Cache injection to zero-latency WASM sandboxing, explore the raw execution mechanics built for unprecedented AI performance.
The Cluaiz Engine provides a full suite of low-level inference controls, bypassing standard REST wrappers to interact directly with shared memory and kernel states.
Execute logic mid-inference. No Python. No HTTP.
The Cluaiz Expression Language (CEL) Router parses raw scripts into an AST that hooks directly into C-Pointers in shared memory. Instead of relying on prompt-engineering for JSON outputs, LogitSteer applies active Token Masking at the logit level — preventing the model from ever generating incorrect syntax. Chain logic, run loops, and conditionals at native CPU speeds with zero network overhead.
Tool results injected directly into attention matrix.
When a tool or WASM skill returns data, Cluaiz does not re-inject it as new context tokens (which would force a full attention re-computation). Instead, it pauses the agentic state at the exact transformer layer, runs the plugin, and uses a C-Pointer to inject the filtered result directly into the active C++ KV-Cache buffer. Eliminates up to 40% context re-computation overhead.
Interrupt. Pivot. Resume. No re-prompt.
Press Ctrl+C during a long generation and Cluaiz pauses the engine state — it does not kill the process. Enter a midway instruction (e.g. 'Make it shorter and add code examples') and the engine processes this pivot via CEL, then resumes generation seamlessly. No context is lost, no re-prompt is required. This is JIT Hot-Steering.
llama.cpp + ONNX Runtime in one unified memory space.
Cluaiz natively orchestrates llama.cpp (for LLM reasoning) and ONNX Runtime (for vision, embeddings, and audio) within a single shared C-level memory space. Hybrid CPU/GPU split-loading enables dynamic context switching between text and vision operations without losing latent states. No duplicate VRAM overhead.
Native capabilities in a 64KB isolated ring.
WASM Skills execute as secure WebAssembly binaries in a strict 64KB isolated sandbox — giving agents native system capabilities (file I/O, OS calls) without REST API round-trip latency. Skills are sandboxed from host memory, preventing malicious plugins from causing Segmentation Faults.
Update YAML manifests without restarting the engine.
Dynamically update YAML manifests — API keys, permissions, model settings, booster profiles — without restarting the inference engine. The Hot-Reload Settings Controller monitors manifest files for changes and applies delta patches atomically. Zero downtime configuration changes during active inference sessions.
