Viswa Bharath Kolla1, Surya Teja Kalavagunta2, Mani Madhuri Manthina 3, 1Independent Researcher, Apex, NC, USA 2Independent Researcher, Rosharon, TX , USA 3Independent Researcher, Rosharon, TX , USA
Retrieval-Augmented Generation (RAG) has emerged as a critical architecture for enterprise AI agents by enabling contextual responses grounded in external knowledge sources. However, in Oracle Cloud SaaS environments, RAG systems suf er from a fundamental limitation: knowledge aging. The issue is primarily attributed to the decoupled nature of data ingestion and static embedding generation, which introduce latency between source data updates and their incorporation into the retrieval index. As enterprise data evolves rapidly, this lag leads to knowledge obsolescence and limits the system’s ability to provide real-time access. This paper presents a dynamic RAG framework using event-driven architecture to bridge the technical debt of knowledge aging and ensure real-time synchronization between transactional systems and AI knowledge layers. A formal model for staleness latency is introduced, along with an ef icient incremental embedding update algorithm. Analytical evaluation demonstrates reduced latency from 𝑂(𝑛)batch updates to 𝑂(𝑘𝑙𝑜𝑔𝑛)incremental updates, where 𝑘≪𝑛.
Retrieval-Augmented Generation(RAG), Oracle SaaS AI Agents, OCI Streaming, Dynamic Embeddings, Event-Driven System
Jiajun Zhang1,Andrew Park2 1Irvine High School, 4321 Walnut Ave, Irvine, CA 92604 2University of California, Irvine, Irvine, CA 92697
Urbanization replaces permeable ground with impervious surfaces, concentrating rainfall into polluted runoff that drives both flooding and water-quality decline, a problem China’s Sponge City program addresses through lowimpact-development (LID) facilities such as bioretention cells, swales, and constructed wetlands. This project proposes a browser-based, real-time 3D simulator that couples a tick-based hydrology and pollutant mass-balance model — tracking TSS, COD, total nitrogen, and total phosphorus — to an interactive isometric city built with Three.js and TypeScript, letting a user toggle LID facilities and rainfall/soil parameters and immediately observe the resulting change in runoff and removal efficiency. Key implementation challenges included preventing numerical instability in the tick loop, correctly chaining serial LID capture without double-counting, and sustaining frame rate under multiple concurrent animated systems. Experiments confirmed that disabling all LID facilities collapses removal efficiency to zero while enabling them restores 72-93% removal, and that rising rainfall intensity overwhelms fixed facility capacity, driving TSS removal from a 96% peak down to 25%.
Sponge city, Low-impact development, Stormwater management, Bioretention, Constructed wetlands, Hydrological simulation
David Sasunov1, Ang Li2, 1Brooks School, 1160 Great Pond Rd, North Andover, MA 01845 2California State University Long Beach, 1250 Bellflower Blvd, Long Beach, CA 90840
Two of the world’s oldest surviving religious traditions independently arrived at the same strange solution to the problem of sacred speech: neither would let scripture simply be spoken. Jewish communities chant the Hebrew Bible according to a fixed system of musical accents, the taamei hamikra, while Buddhist communities across Asia recite their canonical texts according to fixed tonal and rhythmic patterns collectively called chant. This paper argues that both traditions converge on “regular,” patterned sound as a technology for stabilizing the sacred, but that they build this technology on opposite theological foundations and radically different musical logics. Musicologically, both systems exemplify “language as music” — pitch and rhythm fused to a fixed text — yet they diverge sharply in both scale structure and oral-transmission method: Jewish cantillation is built on a small family of richly ornamented modes and generated from the grammar of each verse, while Buddhist chant ranges from the near-monotone reciting tone of Theravada Pali chant to the pentatonic scales of Japanese shomyo, generated from reusable formulas rather than sentence-level syntax. Religiously, Jewish cantillation renders sound the vessel of law, in which correct performance is inseparable from correct meaning, while Buddhist chant renders sound a tool for purifying and concentrating the mind, in which the phonetic act can matter more than semantic comprehension. Anthropologically, both practices function, in Christopher Small’s sense, as “musicking” — communal performances that bind dispersed, linguistically diverse populations into a single group memory, in Maurice Halbwachs’s sense of the term. This paper concludes that “regular sound” recurs across civilizations because it solves three problems simultaneously that ordinary speech cannot: it guarantees the fidelity of oral transmission across generations, it marks a register of speech as set apart from the everyday, and it requires bodies to gather and voice together, manufacturing the very community that the sacred text is meant to define
Sacred Music, Jewish Cantillation, Buddhist Chant, Oral Tradition
Wanqing Tao1, Bobby Nguyen2, 1Fairmont Prep, 2200 W Sequoia Ave, Anaheim, CA 92801 2Cypress College, UC Riverside, 9200 Valley View Street Cypress, CA 90630
Adolescents increasingly bring important life decisions to general-purpose AI chatbots that offer generic advice with no historical grounding, no diversity of perspective, and inconsistent safety behavior for a minor-facing context. This project presents SageCircle, a Flutter mobile application that conditions a large language model on hand-curated biographies of 77 historical figures so that users can discuss real dilemmas with historically anchored “sages,” compare perspectives across a four-figure “Council of Sages,” and journal reflections entirely on-device [3]. Every user message passes through a fail-open moderation gate before reaching the model. Two experiments evaluated the resulting blind spots: a 160-message moderation test found strong detection of direct self-harm language but a recall gap on indirect, teen-voiced phrasing, and a 360-response persona test found that two of six sages exceeded their instructed reply to length on emotionally complex prompts [4]. Compared against persona-conditioned dialogue, scripted therapy chatbots, and memory-based generative agents, SageCircle trades infrastructure weight for biographical grounding and multi-perspective, mobile-friendly reasoning
Large Language Models, Conversational Agents, Persona-Based Dialogue, Mobile Application Development,Content Moderation
Jinrui Jack Fang1,Austin Amakye Ansah2 1Corona Del Mar High School, 2101 Eastbluff Dr, Newport Beach, CA 92660 2The University of Texas at Arlington, 701 S Nedderman Dr, Arlington, TX 76019
Food pantries and community food banks operate under chronic uncertainty: demand fluctuates by weekday and season, incoming donations arrive on irregular schedules, and perishable stock must be moved before it expires [1]. Small organizations rarely have the tooling to anticipate these swings, and the public frequently lacks an accurate, up-to-date directory of where to find help. This paper presents FeedForward, a platform that unifies four capabilities behind role-aware interfaces: a public directory for discovering and locating food resources, an operational inventory tracker for pantry staff, team and branch administration, and a short-horizon demand and supply forecast for registered branches. FeedForward is delivered as two clients over one shared backend: a Flutter mobile application for field staff and the public, and a web portal for administrators and owners [2]. The backend uses Cloud Firestore for real-time data, Firebase Authentication with a four-tier role model, and Cloud Functions for privileged operations [3]. Forecasting runs as a serverless pipeline that combines a seasonal-naive demand model with a deterministic inventory ledger, projecting a seven-day outlook of meals served, supply on hand, and the resulting gap. A parallel Python research stack documents the same approach, adds a Holt-Winters model, and provides a backtesting harness. Evaluated on ten years of public New York City soup-kitchen data, the seasonal-naive model attains a seven-day holdout mean absolute percentage error of 7.7 percent, and Holt-Winters improves this to as low as 5.5 percent over a four-week horizon. To avoid misleading users, forecasts remain in a transparent calibrating state until a branch has at least seven days of history. We describe the architecture, the security model, and the design decisions that keep the platform offline-tolerant, inexpensive to operate, and honest about the limits of its predictions.
Food bank logistics, Demand forecasting, Time-series forecasting, Flutter, Real-time, Mobile computing
Yichen Li1,Andrew Park2 1Crean Lutheran High School, 12500 Sand Canyon Avenue, Irvine, CA 92618 2University of California, Irvine, Irvine, CA 92697
Elderly people living alone face two silent risks: sudden medical emergencies and invisible home hazards, where the delay before help arrives largely determines the outcome. GuardianNest is a single wall-mounted device that fuses mmWave radar, computer vision, and environmental sensors to detect falls, immobility, stroke, loss of consciousness, and gas, smoke, and carbon-monoxide hazards in real time, pushing alerts to a Flutter companion app so families can respond within seconds. A Python detection engine performs sensor fusion with per-event cooldowns, and a Firebase Realtime Database backend streams telemetry while sending push notifications directly from the device — running entirely on a free cloud tier with pairing validated by declarative security rules. Challenges of false alarms, free tier write limits, and serverless secure pairing were addressed in the design. Experiments showed 90.5% average event-detection accuracy and 3.35-second average alert latency. By unifying health and environmental safety in one passive, private, low-cost device, GuardianNest gives caregivers continuous assurance.
Elderly care, Fall detection, mmWave radar, Sensor fusion, Firebase Realtime Database
Yi-chen Cheng1, Austin Amakye Ansah2, 1aipei European School, 31 JianYe Road, Yang Ming Shan, Shihlin District, Taipei, 11193 2The University of Texas at Arlington, 701 S Nedderman Dr, Arlington, TX 76019
Rear-end collisions represent a significant category of vehicular accidents, particularly during parking and low-speed maneuvers where driver visibility is limited. This paper presents StreetRadar, an end-to-end embedded system that combines a 24 GHz mmWave radar sensor (HLK-LD2410C or HLK-LD2420) with a Bluetooth Low Energy bridge running on an ESP32-S3 microcontroller and a companion mobile application built with Flutter [5]. The ESP32 firmware reads radar telemetry over UART at up to 256 kbps, parses both basic and engineering-mode frames, serializes the data into compact JSON payloads, and streams them via BLE notifications at 2 Hz. The mobile app discovers and connects to the radar unit through a custom GATT service, renders a real-time top-down lane visualization with distance-gated color zones, displays per-gate moving and stationary energy profiles, and generates severity-graded proximity alerts [6]. The system operates entirely on-device with no cloud dependency, achieves sub- meter distance resolution over a 3.75–6.0 m range, and supports two sensor models with different gate counts. We describe the firmware architecture, the BLE notification design that avoids common MTU pitfalls, and the cross- platform mobile UI [15].
Vehicular safety, Rear-proximity radar, ESP32-S3, Bluetooth Low Energy, mmWave sensor, Flutter, Embedded systems, Mobile computing
Aaron Zheng1, Rodrigo Onate2, 1St. Paul's School, 325 Pleasant St, Concord, NH 03301 2California State University, Fullerton, 800 N State College Blvd, Fullerton, CA 92831
Veterans transitioning out of military service face compounding physical injuries and psychological challenges, yet existing support tools typically address fitness, mental health, and peer connection as separate, disconnected products. VetStrong proposes a single Flutter and Firebase mobile application that unifies an adaptive fitness engine, an AI- guided wellness companion, a moderated peer community, and a VA resource directory behind one onboarding health survey, so that a veteran’s physical and psychological profile informs every module they use. Key technical components include a Firestore-transaction-based streak tracker, a decoupled chat-service architecture built to accept a generative-AI backend, and a whole-word regular-expression content moderation filter. The central engineering challenge addressed was ensuring the community filter did not produce false positives from embedded substrings, which was resolved through word-boundary tokenization. Two experiments validated this design: the whole-word filter achieved 100% precision, recall, and accuracy against a 180-message test corpus, versus 73.5% precision for a naive substring approach, and an analysis of the 27-exercise catalog confirmed an appropriate lean toward low-impact, accessible movements. VetStrong shows that combining these evidence-backed intervention types into one trusted space is both technically feasible and measurably safer than simpler alternatives.
Veteran health, mobile health (mHealth), PTSD, AI chatbot, Flutter, Firebase
Mylene Zhang1, Andrew Park2, 1St. Margaret's Episcopal School, 31641 La Novia Ave, San Juan Capistrano, CA 92675 2University of California, Irvine, Irvine, CA 92697
Tivvy is a standalone, offline, voice-activated productivity timer designed to help users manage everyday tasks without relying on a smartphone or cloud-connected assistant. The project addresses problems with distraction, convenience, and privacy by using dedicated hardware that processes command locally. Tivvy combines a Nicla Voice chip for keyword spotting, a Qualia S3 RGB666 board for display and timer logic, Bluetooth communication between the boards, speaker feedback, animated visual elements, and a compact enclosure. The system allows users to set, add to, subtract from, cancel, and stop named timers through voice commands while viewing up to three active timers on screen. Major challenges included training reliable voice recognition, fitting the hardware into a practical case, and converting sound effects into a format that the speaker system could use. Planned experiments test command accuracy across noisy environments and battery performance across temperature and usage states. Overall, Tivvy offers a focused, private, and accessible alternative to phone-based timers.
Timer, Voice Activation, Voice Recognition, Machine Learning
