Thesis Defense: Validating XAI, Gamification, and Local-First Architecture

Today marked the culmination of two years of research, development, and iteration—I successfully defended my Master's thesis on NestShield: An AI-Powered Cybersecurity Assistant for Non-Technical Users before my committee at Full Sail University.

Research Validation

The defense validated three core hypotheses through a pre-post intervention study with 42 participants (N=25 completers):

The Defense Experience

The 60-minute defense included a 10-minute live demonstration of NestShield's core features—AEGIS AI explanations, gamification mechanics, and the Threat Radar visualization. The committee posed challenging questions about research validity, technical architecture decisions, and commercialization feasibility. Each question reinforced the importance of acknowledging limitations while demonstrating confidence in the validated findings.

Key Findings

• H1 - Knowledge: XAI-driven education significantly improved security comprehension, with the largest gains in complex concepts like ransomware (33.3% → 87.5%).
• H2 - Behavior: Gamification mechanics achieved a 75% voluntary remediation rate, though a motivation-behavior gap was discovered in 2 participants who felt motivated but lacked enablement.
• H3 - Trust: Local-first architecture built exceptional user trust (6.4/7), with 93% preferring local processing over cloud alternatives. Trust increased 33% after experiencing the system.

Technical Achievements

NestShield integrated cutting-edge technologies into a cohesive educational platform:

• AEGIS with RAG: Mistral 7B LLM powered by Retrieval-Augmented Generation using ChromaDB for contextual, personalized security explanations.
• FastAPI Backend: Async architecture supporting WebSocket connections for real-time scan updates and AI streaming responses.
• React + TypeScript Frontend: Type-safe UI with gamification elements (XP, ranks, Digital Health Score) and interactive threat visualizations.
• 100% Local Processing: Zero cloud dependencies—all AI inference, scanning, and data storage happens on-device, eliminating the "privacy tax."

Retrospective

Engineering Intelligence: Performance Profiling & Adaptive AI

This week, NestShield evolved from a static dashboard into a reactive, intelligent system. The goal was to implement an "Autonomous Consultant" persona—an AI that doesn't just answer questions but analyzes system telemetry to propose remediation steps proactively.

The Challenge: CPU Bottlenecks

While integrating the "Always-On" agent, I encountered a critical performance regression. The global state listener, intended to sync the Dashboard risks with the Header notifications, triggered a race condition in the React rendering cycle. This caused an infinite loop, spiking the browser's Main Thread CPU usage to 100%.

The Solution: Ref-Based Latching

To resolve this, I re-engineered the Chatbot.tsx component using a Reference Latch architecture. By utilizing useRef to track message states outside the render cycle and implementing strict dependency management in useEffect, I eliminated the redundant reconciliations.

Fig 1. Chrome DevTools confirming CPU idle state after optimization.

Retrospective

From Concept to Execution: Profiling NestShield's Knowledge Engine

This week marked a pivotal shift from architectural planning to tangible backend implementation. The focus was Knowledge Ingestion—the process of teaching AEGIS (the AI Brain) to understand complex cybersecurity standards like NIST 800-53 and OWASP Top 10.

The Build

I engineered a Python-based ingestion pipeline (ingest_knowledge.py) utilizing LangChain concepts and ChromaDB for vector storage. The script performs four key actions:

• Parsing: Extracts text from dense, unstructured PDFs using pypdf.
• Chunking: Splits text into 1000-character overlapping segments to preserve context.
• Embedding: Converts text into vector embeddings using a local Transformer model.
• Storage: Indexes vectors into a persistent database for RAG (Retrieval-Augmented Generation).

Performance Profiling Results

A critical requirement for this milestone was performance testing. Using psutil integration and OS-level monitoring, I stress-tested the ingestion of over 600 pages of technical documentation.

The profiling revealed that the operation is heavily Memory-Bound. The significant RAM spike (~733 MB delta) confirms that loading the Transformer model and performing matrix multiplications for embeddings is the primary resource cost. CPU usage spiked in short bursts, which is typical for vectorization tasks.

Fig 1. OS-Level CPU Activity

Fig 2. Script Performance Output

Retrospective

Laying the Foundation: Research & Architecture

The path to building robust software begins long before the first line of code is written. This month, in the Research, Planning, and Design phase of my Master's Capstone, I focused on transforming NestShield from a concept into a tangible architectural plan.

Features & Architecture

My primary goal this month was to design the collaborative intelligence framework. I utilized Python for the backend logic planning and created high-level diagrams to map out the user flow. The biggest challenge was determining how to handle data privacy within a collaborative environment.

Fig 3. Initial Architecture Flow

Retrospective