Wei Jiang 1 Junrui Li 2 , Jiwei Xu 2 , 1 National Police University for Criminal Justice, Baoding, China, 2Beijing University of Chemical Technology, Beijing, China.
Achieving high-efficiency video compression while preserving perceptual quality is a persistent challenge, especially with the growing demand for high-resolution and high-frame-rate content. In this paper, we propose a novel compression framework that focuses on region-of-interest awareness and per-ceptual optimization to enhance both compression efficiency and visual fidelity. The core of our method lies in identifying foreground regions—treated as regions of interest—using an object detection algorithm prior to encoding. These regions receive prioritized treatment during compression, with finer quantization control guided by a deep learning model. This model, based on a convolutional neural network, dynamically predicts quantization levels for each coding block by incorporating both spatial features and perceptual quality metrics. To further improve encoding performance, we reformulate traditional rate-distortion op-timization by introducing data-driven models that relate bitrate and visual quality to quantization levels. These models serve to generate high-quality training labels and guide quantization decisions during in-ference. Additionally, region-aware encoding control adapts quantization granularity based on the size and significance of detected objects. Experimental results demonstrate that the proposed approach signif-icantly reduces bitrate—achieving an maximum saving of around 19%—while maintaining stable and high perceptual video quality, outperforming conventional video coding techniques and recent learning-based methods.
Stephen Bossmart, Marymount University, United States of America
With hybrid work models and multi-cloud environments, perimeter-less networks are making traditional perimeter-based security models obsolete against more complex cyber threats. Ransomware attacks increased by 50% in 2025, and the number of data breaches reached a record high with losses surpassing $20.8 billion. To overcome these challenges, Zero Trust Architecture (ZTA) relies on continuous verification and least-privilege access, assisted by the use of innovative secure protocols such as TLS 1.3, QUIC, and post-quantum cryptography. The study explores the theoretical principles, deployment strategies, and prospects of Zero Trust, highlighting the essence of continuous authentication, micro-segmentation, and AI-driven threat detection. ZTA, combined with next generation secure protocols, is a significant paradigm shift from trust but verify to never trust, always verify, and heightened protection against a growing cyber threat.
Zero Trust Architecture, Secure Protocols, Network Security, Post-Quantum Cryptography, multi factor authentication
Ryder Wei 1 Douglas Winegarden 2 , 1 Cate School, Carpinteria, CA 93013, 2University of Washington Bothell, 17927 113th Ave NE, Bothell, WA 98011
Autistic and neurodiverse learners frequently experience sensory overload in noisy, multi-speaker environments such as classrooms, where an inability to suppress irrelevant auditory input impedes comprehension and participation. This paper presents Voxent, a cross-platform mobile system that translates the selective permeability of a biological membrane into a computational pipeline admitting only the most relevant spoken content. Voxent captures or imports audio, transcribes it with a large-scale weakly supervised speech-recognition model, and reshapes the raw transcript into concise, structure-appropriate notes through a context-conditioned large language model. Five interaction modes — Classroom, Group Collaboration, Family Conversation, Task Sequence, and Public Setting — tailor the output to distinct real-world scenarios. The system is a Flutter client backed by a serverless Firebase architecture in which all model inference runs inside authenticated Cloud Functions, keeping credentials off the device. In a pilot study with twenty-four participants, Voxent earned a System Usability Scale score of 82.4, and 92% of participants preferred mode-formatted notes over raw verbatim transcripts..
assistive technology, automatic speech recognition, neurodiversity, selective attention, large language models, mobile computing, accessibility, cognitive load, serverless architecture, human-computer interaction
Henry Yang 1 Jonathan Sahagun 2 , 1 Sage Hill School, 20402 Newport Coast Dr, Newport Coast, CA 92657, USA , 2 California State University, Los Angeles, 5151 State University Dr, Los Angeles, CA 90032
Inconsistent tooth-brushing contributes to oral diseases that affect roughly half of the global population, yet the behaviour itself is private, repetitive, and unmeasured, which blunts the effect of conventional reminders and verbal instruction. This paper presents Brush Sense, a gamified Internet-of-Things system that couples an accelerometer- equipped smart brush with a mobile application and a cloud backend to measure, score, and socially reinforce daily brushing. An ESP32-S3 device carrying an MPU6050 inertial sensor classifies motion with dual acceleration thresholds, scores vigorous strokes in real time through a combo-and-zone game loop, and streams a self-contained state protocol over Bluetooth Low Energy to a Flutter application backed by Firebase authentication and a Realtime Database holding statistics, streaks, achievements, friends, and leaderboards. Evaluation on the physical prototype shows 97.5% precision and 98.8% recall for session-start detection, a monotonic 2.2× score response across four controlled brushing intensities, and state delivery above a 95% reliability target through five metres.
Smart toothbrush, Internet of Things, gamification, Bluetooth Low Energy, accelerometer, inertial sensing, mobile health, habit formation, embedded systems, Flutter, Firebase, behaviour change
Crystal Zhu 1, Jonathan Sahagun 2 , 1 The Bear Creek School, 8525 208th Ave NE, Redmond, WA 98053, USA , 2 California State University, Los Angeles, 5151 State University Dr, Los Angeles, CA 90032
Children with attention-deficit/hyperactivity disorder and autism spectrum disorder often signal rising dysregulation through repetitive motor behavior such as fidgeting, yet one teacher supervising a dozen students cannot watch every child continuously for these subtle cues. This paper presents WonderWatch, a low-cost wristworn system that classifies a child’s movement into calm, low-fidgeting, or high-fidgeting states in real time and reports the result to a teacher’s phone. The watch pairs an ESP32-S3 microcontroller with a six-axis inertial measurement unit, runs a k-nearest-neighbors classifier entirely on-device, and broadcasts one compact label every 2.5 seconds over Bluetooth Low Energy. A companion Flutter application connects to many watches at once, parses their prediction streams, and displays each child’s live state and accumulated fidgeting time, with rosters stored in Firebase. Under session-aware cross-validation on 176 windows from 14 sessions, the classifier attained a mean accuracy of 0.86, and a hyperparameter sweep confirmed the deployed neighbor count of five sits on a stable plateau.
Wearable sensing, fidget detection, k-nearest neighbors, on-device machine learning, TinyML, Bluetooth Low Energy, inertial measurement unit, special education, edge inference, Flutter, session-aware cross-validation
Dongyue Liu 1, Cesar Magana 2 , 1 Portola high school, 1001 Cadence, Irvine, CA 92618 , 2 California State University Long Beach, 1250 Bellflower Blvd, Long Beach, CA 90840
Small creators frequently encounter significant inefficiencies when navigating between various platforms to monitor performance and obtain critical content feedback. ViewCast resolves these challenges by consolidating essential creator resources into a single, cohesive web interface. This application leverages Firebase Authentication and Hosting, the YouTube Analytics API, and Gemini artificial intelligence [1]. Its core functionality revolves around three vital systems: a comprehensive analytics dashboard, an automated video analyzer, and a predictive thumbnail evaluator. Working in tandem, these components visualize channel metrics, assess video assets, score thumbnail effectiveness, and provide clear, actionable suggestions for improvement. Technical hurdles included maintaining data clarity, designing intuitive navigation, securing sensitive user profiles, and implementing fallbacks for service interruptions [2]. I addressed these by using modular dashboard elements, responsive UI layouts, strict authentication rules, and robust error handling. To validate the system, I performed an experiment using twenty thumbnails to compare predicted quality against actual results. The analyzer achieved an 80 percent success rate by correctly classifying sixteen out of twenty test cases. ViewCast empowers creators with centralized, practical insights that streamline the content optimization process.
Creator Analytics, Artificial Intelligence, Content Optimization, YouTube Analytics
Charlie Zhou 1, Jonathan Thamrun 2 , 1 Bonita High School, 3102 D St, La Verne, CA, USA , 2 University of California, Irvine, Irvine, CA 92697
Retail investors increasingly make decisions inside mobile applications that reduce a model’s output to a single opaque number, overstating confidence and hiding risk. This paper presents Buzzboard, a mobile decision-support system that scores three independent evidence layers—per-headline news sentiment produced by a large language model across CNBC, Bloomberg, and Reuters, social sentiment from StockTwits bullish and bearish message tags, and a rule-based technical-context score—fuses them by reliability weight, and gates the visible forecast on a fourrule signal-strength accumulator that returns “predicted equals current” when the evidence is thin. Every estimate carries a ±1σ band scaled by the square root of the horizon. The system ships as a Flutter client over a Python Flask service with SQLite caching and Firebase persistence. Two experiments on a live Tesla snapshot found the band calibrated to within 0.0025 percentage points of the volatility null, but the point-estimate drift nearly flat across horizons where its own documented formula requires square-root growth, a gap reaching $1.28 at one week.
financial sentiment analysis, large language models, multi-source fusion, realized volatility, forecast calibration, mobile fintech, explainable AI, signal gating, StockTwits, retail investors, Flutter, uncertainty communication
Rasam Emad Yousef Mohammed , School of Computer Science and Technology, Harbin Institute of Technology , Master's Student, Computer Technology
Large language model (LLM)-based multi-agent systems can coordinate through natural-language communication, shared tasks, and autonomous actions, creating risks that may not be visible from final task performance alone. This paper proposes a two-part framework for detecting collusion in such systems. Explicit collusion is detected from communication evidence, including collusive intention, coalition formation, persuasion or pressure, concealment, and consistency between messages and actions. Implicit collusion is assessed from behavioral outcomes, including performance loss, regret, repeated coalition advantage, and inconsistency between apparently cooperative communication and harmful actions. The proposed evaluation combines structured LLM-based judgment with outcome-based analysis under controlled repeated experiments. The framework is designed to distinguish normal coordination, attempted collusion, explicit collusion, and hidden coordination while reducing false positives through baselines and repeated observations.
LLM-based multi-agent systems, collusion detection, explicit collusion, implicit collusion, coalition advantage
Yundi Xu 1, Austin Amakye Ansah 2 , 1 Woodside Priory School, 302 Portola Valley, Irvine, CA 94028 , 2 The University of Texas at Arlington, 701 S Nedderman Dr, Arlington, TX 76019
Hearing loss is a growing concern among young adults due to increased exposure to personal audio devices, yet access to routine clinical testing remains inconvenient and expensive [1]. To address this, we propose HearWell, a mobile application designed for accessible hearing health screening, data management, and trend analysis. Because standardizing absolute hearing thresholds across uncalibrated consumer devices is highly challenging, HearWell employs a modified Hughson-Westlake procedure to measure relative hearing performance across frequencies, using 1000 Hz as a baseline. Built with Flutter, Firebase, and a Flask backend, the app features in-app pure-tone and tinnitus screenings, secure audiogram storage, and retrieval-augmented generation (RAG) AI summaries [2][3]. By focusing on frequency-specific weaknesses and performance trends rather than absolute clinical diagnostics, HearWell empowers users to continuously monitor their hearing health. Furthermore, integrated appointment scheduling and educational resources promote proactive long-term hearing care and early intervention.
Hearing, Audiology, Hughson-Westlake, Tinnitus
Dongyue Liu 1, Cesar Magana 2 , 1 Portola high school, 1001 Cadence, Irvine, CA 92618 , 2 California State University Long Beach, 1250 Bellflower Blvd, Long Beach, CA 90840
Small creators frequently encounter significant inefficiencies when navigating between various platforms to monitor performance and obtain critical content feedback. ViewCast resolves these challenges by consolidating essential creator resources into a single, cohesive web interface. This application leverages Firebase Authentication and Hosting, the YouTube Analytics API, and Gemini artificial intelligence [1]. Its core functionality revolves around three vital systems: a comprehensive analytics dashboard, an automated video analyzer, and a predictive thumbnail evaluator. Working in tandem, these components visualize channel metrics, assess video assets, score thumbnail effectiveness, and provide clear, actionable suggestions for improvement. Technical hurdles included maintaining data clarity, designing intuitive navigation, securing sensitive user profiles, and implementing fallbacks for service interruptions [2]. I addressed these by using modular dashboard elements, responsive UI layouts, strict authentication rules, and robust error handling. To validate the system, I performed an experiment using twenty thumbnails to compare predicted quality against actual results. The analyzer achieved an 80 percent success rate by correctly classifying sixteen out of twenty test cases. ViewCast empowers creators with centralized, practical insights that streamline the content optimization process.
Creator Analytics, Artificial Intelligence, Content Optimization, YouTube Analytics
Leo Tao 1, Austin Amakye Ansah 2 , 1 Tesoro High School, Tesoro Creek Rd, Rancho Santa Margarita, CA 92688 , 2 The University of Texas at Arlington, 701 S Nedderman Dr, Arlington, TX 76019
Rink Rivals is a real-time 3D arcade ice-hockey shooting gallery that explores natural, embodied input as an alternative to gamepad and keyboard control. It is built on Godot 4.7 with Jolt rigid-body physics. In the primary mode, Virtual Ice Challenge, the player stays stationary and aims a limited number of pucks at a rule-based goalie using a webcam and on-device MediaPipe models: a pose landmarker maps either raised hand to screen-space net aim, and a gesture recognizer confirms ready-pose calibration with a thumbs-up [5]. Swinging the play hand transfers measured hand speed into puck flight speed, while aim direction follows the pointed screen target. Keyboard and gamepad remain available as secondary modalities behind the same player controller. Supporting systems include dual-hand soft-pause warnings, a first-person stick with procedural swing IK, an arm-preview viewport, arcade puck containment, and a Beehave behavior tree that positions the goalie [4]. This paper documents the architecture and design decisions behind each subsystem as realized in the current codebase.
Pose Estimation, MediaPipe, Gesture Recognition, Behavior Trees, Godot Engine
Chenkai Xu 1, Moddwyn Andaya 2 , 1 Savannah College of Art and Design, 516 Drayton St, Savannah, GA 31401 , 2 California State University, Sacramento, 6000 Jed Smith Dr, Sacramento, CA 95819
Many players value the long-term growth that role playing games offer, but conventional systems assume large blocks of continuous attention that busy schedules rarely allow. This project addresses that gap with an idle progression game, built in Unity, in which job and skill progression, an item economy, and automated combat continue advancing on their own timers so that meaningful choices, not constant input, drive progress [7]. The system is organized around a small number of interconnected components managed through centralized, singleton-based logic and a shared data catalog, keeping the structure easy to extend. Anticipated challenges included pacing skill progression evenly, preventing the item economy from becoming unbalanced, and keeping automated combat fair, each addressed through shared, formula driven systems rather than hand tuned values. Testing combat across a wide range of monster levels showed the system holding up reliably within its intended range before revealing a sharp balance cliff beyond it, a concrete and useful finding for future tuning. Overall, the project demonstrates that meaningful, long term game progression does not require the continuous engagement that conventional designs assume.
Idle Games, Incremental Games, Game Progression Systems, Unity Game Development
Copyright © CLSB 2026
