AI-Driven Adaptive Tutoring: A Multi-Agent System for Structured, Multimedia-Enhanced Education

Type: Undergraduate
Author(s): Jayapradeep Jayaraman Srinivas Computer Science Gabriella Campos Computer Science Robin chataut Computer Science Nagendra Chaudhary Computer Science Riley Phan Computer Science
Advisor(s): Robin Chataut Computer Science

We present the AI-Driven Adaptive Tutoring (AIAT) framework, a modular multi-agent system that generates structured, retrieval-grounded, and multimedia-enhanced courses. AIAT targets a common gap in AI in Education: large language models (LLMs) can produce fluent explanations, but they often lack pedagogical structure, factual grounding, and multimodal integration. To address this, AIAT uses a three-stage pipeline. First, a blueprint agent creates a course outline with learning objectives and topic dependencies using schema-validated structured outputs. Second, a chapter-expansion agent instantiates atomic topics with formative questions and summaries in JSON mode. Third, an enrichment agent generates topic-level explanations, visualization specifications, and triggers for narrated video production. Retrieval-augmented generation (RAG) combines a MongoDB Atlas Vector Search backend for course materials and a Pinecone pipeline for PDF-derived knowledge, grounding explanations in external content. A Next.js frontend streams responses and assembles text, diagrams, and videos into a unified learner experience.

The design is explicitly guided by mastery learning, cognitive load theory, and the Cognitive Theory of Multimedia Learning, with principles such as atomic topics, anti-fluff constraints, and visual-verbal alignment encoded in prompts and schemas. We report system-level metrics (e.g., latency by component) and operational reliability, and we outline a concrete evaluation plan, including pre/post-learning assessments, expert rubric-based accuracy checks, and subjective cognitive load measures. We also discuss the equity and accessibility implications of relying on commercial APIs and propose mitigation strategies (e.g., caching, partial use of lightweight models, and instructor-in-the-loop authoring). The contribution of this work is a reproducible architecture that connects multi-agent orchestration, RAG, and multimodal rendering to pedagogical theory, along with an evaluation roadmap that explicitly addresses the current lack of large-scale human studies.

HelioIndex: A Python Package for Reproducible Construction of Machine-Learning Datasets from Solar Observations and Event Catalogs

Type: Undergraduate
Author(s): Balaji Kannan Computer Science Robin Chataut Computer Science
Advisor(s): Chetraj Pandey Computer Science

Space Weather Forecasting relies on large volumes of time-stamped solar observations paired with event catalogs describing flare occurrence and intensity. While these datasets are widely available, preparing them for machine learning remains a substantial and often overlooked challenge. Researchers must convert irregular observation streams into consistent temporal samples, construct observation and prediction windows, align events with observations, manage missing data and cadence inconsistencies, and ensure that training and evaluation splits avoid temporal or regional data leakage. These preprocessing steps are typically implemented in ad-hoc scripts that are difficult to reproduce, extend, or compare across studies. We propose an open-source Python library that standardizes the construction of machine-learning-ready datasets for solar event forecasting. The library ingests user-provided observation tables (e.g., SDO image timestamps and file paths) and event catalogs (e.g., GOES flare lists) and automatically generates indexed training samples suitable for PyTorch datasets and data loaders. Users can define flexible observation windows ranging from single time points to multi-frame temporal sequences, specify prediction horizons, and configure event-labeling rules. The framework also provides mechanisms for handling missing observations, irregular cadences, and explicit representation of temporal gaps. To support rigorous experimental design, the library includes reproducible dataset partitioning strategies such as chronological and tri-monthly splits, as well as optional active-region-aware grouping based on NOAA region catalogs. These features allow researchers to build both full-disk and active-region-based forecasting datasets while minimizing common sources of information leakage. Although the initial implementation focuses on solar flare prediction, the framework is designed to be extensible to other space weather phenomena including coronal mass ejections (CMEs) and solar energetic particle (SEP) events. By formalizing the transformation from raw observational records and event lists into reproducible machine learning datasets, this tool reduces the overhead of data preparation and promotes more transparent, comparable, and scalable space weather forecasting research.

A Cross-Model Validation Framework for Solar Flare Forecasting Using NASA CCMC Flare Scoreboard Predictions

Type: Undergraduate
Author(s): Kritika Karanjit Computer Science Robin Chataut Computer Science Chetraj Pandey Computer Science
Advisor(s): Chetraj Pandey Computer Science

Solar flares are significant space weather phenomena that can impact satellites, communication systems, and many technological infrastructures, rendering accurate flare forecasting a crucial objective in heliophysics study.The NASA Community Coordinated Modeling Center (CCMC) Flare Scoreboard collects predictions from multiple solar flare forecasting models developed by different research groups. While this resource provides a useful platform for comparing different forecasting approaches, systematic validation of these models remains challenging because predictions are reported in different formats and are not easily comparable across models. In this work, we develop an automated framework to collect and organize flare forecasts from several models available in the CCMC Flare Scoreboard and convert them into a consistent dataset that allows direct comparison between models. The processed dataset includes predictions across multiple years and forecast windows. To evaluate model performance, we compare the predicted flare probabilities with observed flare events reported in the SolarSoft (SSW) Latest Events archive. By aligning forecast windows with actual flare occurrences, we establish a consistent approach for validating model predictions. This approach facilitates a systematic comparison of forecasting behaviour among various models and assists in identifying those that exhibit superior or inferior predicted ability.The resulting pipeline provides a reproducible way to analyze solar flare forecasting systems and supports future efforts to improve the reliability of space weather prediction methods.

Integrating Full-Disk and Active-Region Models for Reliable Forecasting of Solar Flares

Type: Undergraduate
Author(s): Duc Le Computer Science Robin Chataut Computer Science Chetraj Pandey Computer Science
Advisor(s): Chetraj Pandey Computer Science

Solar flares are major drivers of space-weather disturbances and can disrupt satellites, communication systems, and navigation infrastructure. Recent deep learning approaches have demonstrated promising performance for solar flare forecasting, yet many models operate either on full-disk solar observations or on isolated active-region patches. This separation limits their ability to combine global solar context with localized magnetic structure and can affect the reliability of predictions. In addition, full-disk models often provide limited information about which regions drive their forecasts. This study presents a two-stage deep learning framework that integrates full-disk and active-region–level analysis within a unified flare forecasting pipeline. The system first performs full-disk inference using a convolutional neural network trained on solar magnetograms to estimate the global probability of flare occurrence. Attribution-based explanations are then generated to identify regions that most strongly influence the model prediction. These regions are mapped back to the solar disk and converted into candidate active-region patches, accounting for solar rotation and spatial alignment. The resulting patches are subsequently analyzed using a dedicated active-region forecasting model trained on SDO HMI SHARP data to produce localized flare probabilities. By integrating global context with targeted active-region analysis, the proposed framework combines two complementary forecasting models into an end-to-end prediction system. The resulting pipeline provides both full-disk and region-level flare probabilities, improving interpretability while enhancing the reliability of flare forecasts through focused secondary analysis of the most relevant solar regions.

Using AlphaFold2 to Identify Novel Drug Targets Against Cryptococcus

Type: Undergraduate
Author(s): Francisco Lugo Gonzales Computer Science
Advisor(s): Natalia Castro Lopez Biology