Systems Research
KB (Kernos + BNLM)
A browser-native operating environment unifying Kernos OS systems architecture with BNLM client-side machine learning inside a single tab. Features an in-browser intent classifier trained in seconds (~8k parameters) with per-character occlusion attribution, a capability-governed VFS terminal with pipes, sandboxed Pyodide Python WebAssembly execution, and Groq cloud intelligence fallback.
Engineering objective
Deliver a private, zero-install operating environment in the browser capable of training and running private intent classifiers entirely offline with explainable attribution, capability policies, and sandboxed Python execution.
Architecture
KB operates as a full client-side cognitive operating environment. Users can train a private ~8k-parameter intent router on local data in seconds, inspect decisions with per-character occlusion attribution, execute Python scripts in Pyodide WebAssembly, and navigate a capability-governed VFS terminal with zero data leaving the browser tab.
- Local-First Privacy: Intent classification and explainability execute entirely inside the client tab without telemetry or server round-trips.
- Honest Held-Out Re-evaluation: Model corrections are evaluated against frozen held-out test splits to measure authentic generalization.
- Capability-Based Security: Terminal commands and agents operate under explicit capability policies (can / policy built-ins).
- Embedded BNLM Engine: Vendored client-side machine learning engine under src/bnlm/ providing in-browser tensor math, embedding, classification, and WebGPU compute shaders.
Implementation evidence
Per-Character Occlusion Attribution
OBSERVEDMeasuring prediction shift under character occlusion identifies specific token triggers without neural network opacity.
Sub-Second In-Tab Training
OBSERVEDTraining an ~8k parameter classifier in browser memory completes in under 3 seconds, enabling instant interactive correction loops.