NeuroSkill(tm): Proactive Real-Time Agentic System Capable of Modeling Human State of Mind

NeuroSkill™ is a real-time, proactive agentic system capable of modeling the human state of mind by leveraging biophysical and brain signals from Brain-Computer Interface (BCI) devices such as EEG headsets . It operates fully offline on the edge, utilizing a foundation EXG model and text embeddings model to analyze neural and physiological data without requiring cloud connectivity or internet access after installation . The system provides an API and CLI that expose a structured description of the user’s state of mind through a file format called SKILL.md, which encapsulates cognitive, affective, and neurological states .

The system supports hardware including Muse 2, Muse S, and OpenBCI boards (Ganglion, Cyton, Cyton+Daisy, Galea), capturing 4-channel EEG (TP9, AF7, AF8, TP10), PPG (HR, HRV, SpO₂), and 9-axis IMU data . It computes over 70 real-time brain metrics—such as spectral power, complexity, connectivity, and artifact detection—and uses GPU-accelerated transformer encoders (via wgpu) to generate 32-dimensional neural embeddings from 5-second EEG epochs . These embeddings enable vector similarity search using an HNSW index for fast retrieval of similar brain states across recording history .

NeuroSkill™ integrates with NeuroLoop™, an agentic harness that injects live EEG context into AI interactions before each response, enabling empathetic and cognitively aware AI behavior . This context includes real-time cognitive load, relaxation, engagement, drowsiness, mood, and heart rate, updated every second . The system supports automatic sleep staging (Wake, N1, N2, N3, REM) per AASM standards and offers session comparison, label annotation, and guided protocols .

All data remains local—stored in CSV, SQLite, and HNSW binary formats—with no telemetry, accounts, or outbound network requests during normal operation . The software is open-source under GPLv3 with ethically aligned AI100 licensing for skill markdowns, emphasizing privacy, transparency, and user control . It is built using Rust, Svelte 5, wgpu, and Tauri 2, and runs natively on macOS, with potential support for Linux and Windows

NeuroSkill introduces a proactive, real-time agentic framework designed to bridge the gap between physiological telemetry and semantic understanding. By operating as an offline edge system, it processes raw electrophysiological signals (EXG/BCI) alongside text embeddings to construct a dynamic model of the user's state of mind. The system leverages a novel "SKILL.md" descriptor format, which functions as a structured, human-readable configuration layer that maps biological data patterns to cognitive states or actionable intents. This architecture allows the agent to interpret complex neural and physiological inputs without relying on cloud connectivity, ensuring low-latency response and high data privacy—critical factors for wearable bio-feedback systems.

The primary contribution of this work is the integration of bio-signal processing with semantic embeddings within a unified, modular agentic loop. Unlike traditional Brain-Computer Interfaces (BCI) that rely on rigid, pre-trained classification of specific motor commands, NeuroSkill focuses on holistic state modeling—inferring cognitive conditions such as focus, fatigue, or stress. The utilization of SKILL.md descriptions suggests a shift toward declarative AI programming, where the agent's behavior and interpretation logic can be modified or extended through text-based definitions rather than retraining underlying models. This approach facilitates a more flexible and adaptive interaction paradigm between humans and machines.

This research is significant as it represents a step toward truly proactive computing systems that can anticipate user needs based on internal physiological states rather than explicit external commands. By enabling agents to understand and react to the human "state of mind" in real-time, NeuroSkill moves beyond reactive command-and-control interfaces toward anticipatory assistance. The edge-based deployment addresses critical barriers in BCI adoption, specifically privacy concerns and latency issues, making it a viable blueprint for future consumer wearables and assistive technologies.

# Summary: NeuroSkill™ – Proactive Real-Time Agentic System for Human State Modeling

NeuroSkill™ is a novel real-time, offline edge system designed to model a human’s cognitive and emotional state using a combination of electrophysiological signals (EXG/BCI) and text embeddings. The system leverages SKILL.md—a structured, human-readable language for describing mental states—to generate proactive, context-aware interventions without requiring cloud connectivity. By integrating brain-computer interface (BCI) data with natural language processing (NLP) embeddings, NeuroSkill enables real-time state estimation, allowing agents to adapt their behavior dynamically based on inferred mental modes such as focus, stress, or fatigue.

The paper’s key contributions include: 1. A unified framework for fusing neural and textual signals to model human state, enabling low-latency, privacy-preserving inference on edge devices. 2. Proactive agentic behavior, where the system not only detects but also anticipates and mitigates cognitive load or emotional states by adjusting task difficulty, pacing, or providing feedback. 3. Offline capability, ensuring robustness in environments where cloud access is restricted (e.g., military, medical, or remote settings).

This work is significant because it bridges neuroscience, BCI, and AI agent systems, offering a practical, real-world deployment-ready solution for applications in mental health monitoring, adaptive learning, workplace productivity, and human-AI collaboration. By enabling autonomous, context-aware agents that respond to human state in real time, NeuroSkill could redefine how AI systems interact with users in high-stakes, dynamic environments.

Source: [arXiv:2603.03212](https://arxiv.org/abs/2603.03212)