OpenSparX/MasterAgent

What it solves

OAK (Open Agent Kernel) provides a high-performance, 100% on-device kernel for AI agents, eliminating the need for cloud dependencies, reducing latency to sub-100ms, and ensuring data privacy by keeping all processing on the local hardware.

How it works

OAK uses a hybrid routing system where approximately 80% of requests are handled by a deterministic pattern-matching skill engine (responding in microseconds) and the remaining 20% are routed to a local LLM. It features a Task Orchestrator for DAG execution and supports hardware acceleration via the Qualcomm QNN backend for NPUs.

Who it’s for

Developers building privacy-focused, low-latency agents for automotive systems, smart home orchestration, IoT edge devices, and industrial automation.

Highlights

• Speculative Execution: Predicts user intent and pre-computes responses during idle NPU time to achieve near-instant responses.
• Formal Plan Verification: Uses CTL* model checking to verify that agent plans are safe and deadlock-free before execution.
• Agent Mesh Protocol: Enables zero-config multi-device collaboration using mDNS, CRDT state sync, and Merkle anti-entropy to route tasks to the most capable device.
• Crash Recovery (WAL): Implements a Write-Ahead Log with an "UNKNOWN" state to prevent duplicate actions (like double charging) after a system crash.
• On-Device Continual Learning: Supports local QLoRA fine-tuning with Differential Privacy (DP-SGD) to improve the agent based on user corrections without sending data to the cloud.