Ramteja Sajja

The rapid evolution of Mixed Reality (MR) technologies, particularly Virtual Reality (VR), offers powerful new means of visualizing and interacting with geographic and environmental data. This paper presents Virtual Reality Flood Information System (VRFIS), an immersive information platform developed using Unreal Engine 5 and Google Photo Realistic 3D Tiles to enable real-time exploration of high-resolution geospatial datasets across the United States. VRFIS integrates photorealistic terrain rendering, dynamic environmental simulation, and multimodal interaction within a fully immersive 3D environment, allowing users to visualize, query, and interpret complex environmental phenomena with unprecedented realism and spatial context. By coupling high-fidelity visualization with live data streams from authoritative sources such as USGS and NOAA, the system serves as an intuitive visualization and educational tool for environmental science, hydrology, urban planning, and geography education. To assess the system's practical utility and identify usability constraints, a qualitative expert evaluation was conducted with domain specialists, including hydrologists and geospatial analysts. The heuristic feedback demonstrates that while existing 2D web-based platforms excel at quantitative data management, immersive environments like VRFIS uniquely enhance spatial awareness, promote experiential learning, and provide volumetric context that can support environmental decision-making.

This study introduces the Synthetic Crowd Wisdom (SCW), a concept designed to emulate the "wisdom of crowds" mechanism without human participants using large language models. The accuracy, calibration, and robustness gains of the approach is obtained by aggregating responses from independent, persona-diverse artificial intelligence (AI) agents under explicit aggregation rules. We present Crowd Wisdom AI, a multi-agent framework that operationalizes SCW by simulating diverse human perspectives to study and enhance collective intelligence. Beyond a technical platform, it serves as a behavioral research tool for examining how demographic and cognitive diversity shape group reasoning and consensus formation. Each agent embodies distinct traits, such as age, education, cultural background, and problem-solving style, allowing controlled exploration of collective decision dynamics. Aggregated responses reveal consensus patterns that mirror human crowd effects, providing a means to investigate conditions under which AIsimulated groups approximate real-world decision-making. Early experiments highlight strong performance in estimation and visual interpretation tasks. In water-level sensor readings, where cheating is impossible due to objective measurement, the system achieved an error margin of only 0.45 ft (2.9%), closely aligning with actual values. These results suggest that leveraging Synthetic Crowd Wisdom not only improves decision quality but also provides a scalable computational framework for modeling human cognitive diversity, advancing behavioral research into the psychological mechanisms of collective intelligence.

Fine-grained emotion verification remains a persistent challenge for large language models, where supervised classifiers struggle with taxonomic subtlety and class imbalance, plateauing at 0.56 F1 on the 28-class GoEmotions benchmark. We propose that this limitation stems from semantic sparsity in standard token representations and introduce Semantic Density Optimization, a geometric framework that amplifies discriminative signals within high-dimensional embedding spaces through controlled synonym expansion. We implement a training-free verification pipeline that expands sparse sentiment indicating terms into dense, synonym-enriched representations using GPT-5.2-based extraction, validating the approach on outputs from a Qwen3-8B model fine-tuned for emotion control. On these synthetic responses, our method achieves 71.61% accuracy (0.68 F1), outperforming supervised RoBERTa on identical data by 20.7 percentage points. Systematic analysis reveals that increasing semantic density from sparse (2-3 terms, 47.98%) to dense representations (20-30 terms, 71.61%) yields a 23.63 percentage point improvement, enabling the detection of rare emotions like relief and realization that supervised models fail to recognize. These results demonstrate that geometrically optimizing semantic representation offers a superior, zero-shot alternative to supervised learning for fine-grained emotion classification, requiring no task-specific training while providing interpretable explanations.

As generative AI becomes increasingly integrated into higher education, understanding how students engage with these technologies is essential for responsible adoption. This study evaluates the Educational AI Hub, an AI-powered learning framework, implemented in undergraduate civil and environmental engineering courses at a large R1 public university. Using a mixed-methods design combining pre- and post-surveys, system usage logs, and qualitative analysis of students’ AI interactions, the research examines perceptions of trust, ethics, usability, and learning outcomes. Findings show that students valued the AI assistant for its accessibility and comfort, with nearly half reporting greater ease using it than seeking help from instructors or teaching assistants. The tool was most helpful for completing homework and understanding concepts, though views on its instructional quality were mixed. Ethical uncertainty, particularly around institutional policy and academic integrity, emerged as a key barrier to full engagement. Overall, students regarded AI as a supplement rather than a replacement for human instruction. The study highlights the importance of usability, ethical transparency, and faculty guidance in promoting meaningful AI engagement. A total of 71 students participated across two courses, generating over 600 AI interactions and 100 survey responses that provided both quantitative and contextual insights into learning engagement.

Large Language Models (LLMs) have shown strong performance across natural language processing tasks, yet their general-purpose embeddings often fall short in domains with specialized terminology and complex syntax, such as hydrology and environmental science. This study introduces HydroEmbed, a suite of open-source sentence embedding models fine-tuned for four QA formats: multiple-choice (MCQ), true/false (TF), fill-in-the-blank (FITB), and open-ended questions. Models were trained on the HydroLLM Benchmark, a domain-aligned dataset combining textbook and scientific article content. Fine-tuning strategies included MultipleNegativesRankingLoss, CosineSimilarityLoss, and TripletLoss, selected to match each task's semantic structure. Evaluation was conducted on a held-out set of 400 textbook-derived QA pairs, using top-k similarity-based context retrieval and GPT-4o-mini for answer generation. Results show that the fine-tuned models match or exceed performance of strong proprietary and open-source baselines, particularly in FITB and open-ended tasks, where domain alignment significantly improves semantic precision. The MCQ/TF model also achieved competitive accuracy. These findings highlight the value of task- and domain-specific embedding models for building robust retrieval-augmented generation (RAG) pipelines and intelligent QA systems in scientific domains. This work represents a foundational step toward HydroLLM, a domain-specialized language model ecosystem for environmental sciences.