This research investigates the integration of Small Language Models (SLMs) on edge computing hardware to drive autonomous, conversational Non-Player Characters (NPCs) within Virtual Reality (VR) training environments without reliance on cloud infrastructure. By utilizing an NVIDIA Jetson Orin Nano as a localized inference server, the study aims to pipeline Speech-to-Text, SLM response generation, and Text-to-Speech processes while strictly maintaining the low latency required for VR immersion. The findings will establish a framework for deploying privacy-preserving, context-aware conversational agents in decentralized or offline spatial computing applications.

• To configure and containerize a localized AI stack on the Jetson Orin Nano, integrating a highly quantized Small Language Model (e.g., Llama-3.2-1B) alongside local speech recognition and synthesis modules.
• To establish a low-latency, bidirectional network protocol (e.g., WebSockets or MQTT) that links the VR application (built in Unity/Unreal Engine) with the Jetson inference node for real-time data exchange.
• To measure the end-to-end latency (Time-to-First-Token, overall interaction delay) and system stability under load, determining if the response times meet the human-computer interaction thresholds required to maintain immersion in VR.

• A working prototype of a VR training scenario featuring an offline, AI-driven NPC.
• Benchmarking data documenting the memory usage, thermal performance, and response latency of the Jetson device under conversational load.
• A reproducible architectural framework for organizations that need smart VR training but have strict data-privacy or offline requirements (e.g., healthcare, defense, industrial safety).

This research focuses on developing an AI-powered web platform to automate the compilation of academic conference proceedings, tailored to publisher standards such as Springer's. By leveraging Document AI and Large Language Models, the system will ingest heterogeneous submission files (PDFs, LaTeX/Word sources, and license agreements) to automatically extract metadata, reconcile inconsistencies, and dynamically calculate pagination. This approach aims to eliminate manual editorial bottlenecks, automatically generate required structural artifacts, and ensure strict formatting compliance.

• To develop a multi-format ingestion pipeline capable of parsing PDFs, source files, and scanned license agreements to extract structured metadata (names, ORCIDs, emails).
• To integrate an AI-driven reconciliation module that detects and resolves inconsistencies across submission files, ensuring accurate affiliations and corresponding author flags.
• To implement a dynamic compilation engine that automatically generates publisher-compliant artifacts, including a synchronized Table of Contents and an alphabetical Author Index.

• A fully functional web prototype allowing authors and editors to upload, validate, and manage proceedings files in a centralized environment.
• Automated generation of publisher-ready deliverables (Table of Contents, Author Index, Excel metadata lists) with precise, dynamically calculated page numbering.
• Benchmarking data demonstrating a significant reduction in editorial processing time and manual pagination errors compared to traditional workflows.

The proliferation of covert wearables, such as smart glasses and micro-earpieces, poses a critical challenge to academic integrity. Traditional visual proctoring is often insufficient to detect these devices in high-density exam environments. This thesis proposes a passive Radio Frequency (RF) monitoring framework utilizing open-source sniffers (Ubertooth One, CatSniffer) and directional antennas to detect and localize active cheating. By profiling 2.4 GHz traffic signatures, the proposed system distinguishes benign, idle background devices (e.g., smartphones in bags) from the high-throughput data streams characteristic of live audio calls or media transfers (e.g., Meta Ray-Bans). Coupled with RSSI-based spatial triangulation, this framework will provide real-time, accurate identification and localization of unauthorized transmissions during exams, ensuring a secure testing environment without compromising student data privacy.

• Characterize and document the distinct 2.4 GHz traffic signatures, packet types (e.g., SCO/eSCO synchronous data for Classic audio, Isochronous BLE Audio, Wi-Fi Direct bursts), and throughput profiles of target cheating devices (Meta Ray-Bans, covert micro-earpieces, standard wireless earbuds) compared to idle background devices (smartphones in bags, smartwatches).
• Develop a software pipeline/script (leveraging tools like Wireshark/Pyshark, Kismet, or custom Python scripts) that filters out passive BLE advertisements and flags anomalous, high-throughput, sustained data streams in real time.
• Integrate the sniffing hardware with a 2.4 GHz directional antenna (Yagi/patch) to evaluate the spatial accuracy of RSSI-based direction-finding and locate target devices despite human body attenuation (e.g., head shadowing from micro-earpieces).
• Conduct controlled experiments in a simulated exam hall to quantify the system's detection latency, false-positive rate, physical localization error margin (in meters), and performance under high device density.

• Wearable RF Fingerprint Library: A comprehensive behavioral profile detailing how specific devices (Meta Ray-Bans taking photos/streaming audio vs. AirPods on an active cellular call vs. idle phones) behave in the 2.4 GHz spectrum.
• Real-Time Detection Tool: A working software prototype that monitors live packet streams from a Ubertooth or CatSniffer, automatically triggers an alert upon detecting an active two-way audio stream or Wi-Fi Direct video offload, and logs the offending MAC address or connection parameters.
• Spatial Triangulation Evaluation: Documented empirical performance data showing the effective range, angle of arrival (AoA) accuracy, and signal attenuation metrics when sweeping an exam room with a directional antenna.
• A practical, ethical protocol and deployment guide outlining how proctors and educational institutions can legally and non-intrusively monitor RF spectrum without violating student privacy or decrypting personal communications.

This study presents an automated NLP-based framework for analyzing four years of unstructured student course evaluations to support continuous curriculum improvement. Using sentiment analysis and topic modeling techniques (LDA, BERTopic), the system identifies latent themes and correlates sentiment fluctuations with historical curricular and administrative changes. Findings are delivered through an anonymized visualization dashboard designed for course coordinators. Key challenges include handling noisy informal text, mitigating algorithmic bias, and distinguishing constructive feedback from non-actionable responses. Patterns are validated against quantitative performance metrics to ensure reliability, offering institutions a scalable, privacy-preserving tool for evidence-based pedagogical decision-making.

Student