Skip to content

Research Overview

The Computer Science Education (CSED) Lab at UC Merced investigates how artificial intelligence and modern software infrastructure can transform computer science pedagogy. Led by PI Santosh Chandrasekhar and PhD student Michael Pisman, the lab's research sits at the intersection of AI-assisted assessment, computational skill modeling, algorithmic team formation, human-computer interaction, and cloud-native computing. Each thrust is driven by real instructional pain points observed in university CS programs and validated through empirical studies with student populations.

A unifying theme across the lab's projects is the use of formal computational methods to solve ill-defined educational problems. Whether converting natural language course materials into structured assessment items, inferring prerequisite skill relationships from unstructured text, or formulating team composition as an integer linear program, the lab treats CS pedagogy as a modeling and optimization challenge. This approach is grounded in the Model Context Protocol (MCP) as an integration layer, connecting large language models to solvers, databases, and knowledge bases in a principled, auditable way.

Research Thrusts

AI-Assisted Assessment Generation — Test Forge. A system that ingests instructor-provided course materials and uses retrieval-augmented generation (RAG) combined with LLMs to produce assessment questions aligned to Bloom's Taxonomy. Test Forge gives instructors fine-grained control over cognitive difficulty distribution and supports an interactive feedback loop that enables dynamic editing and model refinement, moving beyond black-box test generators toward a collaborative, pedagogically-informed workflow.

Skill Networks and Graph-Based Knowledge Representation — SkillNet. A framework for constructing, validating, and querying skill graphs from heterogeneous data sources. SkillNet uses LLMs with chain-of-thought auditing to extract prerequisite and co-occurrence relationships from unstructured text, seeds the graph with the Computer Science Ontology (CSO), and validates structural quality through small-world network metrics. Graph Neural Networks (GNNs) operate on the constructed graph to infer latent relationships and generate embeddings used for learning-path construction and team formation.

Algorithmic Team Formation — OptiTeam. A natural-language-to-optimization pipeline that converts unstructured instructor requirements into formal integer linear programming (ILP) models for team composition. OptiTeam employs a multi-agent LLM architecture — with manager, formulator, and evaluator agents — to generate, debug, and refine constraints. The system explores the fairness-optimality trade-off in team allocation and integrates with constraint solvers through the Model Context Protocol.

Multi-Modal Attendance Verification — Tendy. A cross-platform attendance system that investigates the fundamental tension between friction, trust, and privacy through three distinct check-in modalities: Dynamic QR with sliding-window TOTP, Bluetooth Low Energy proximity, and biometric liveness detection. A pilot study using the System Usability Scale (SUS) revealed that higher-friction methods (Dynamic QR: SUS 83.9) paradoxically scored significantly higher than invisible passive methods (BLE: SUS 59.6).

Serverless Computing for Educational Infrastructure. A comparison of deployment models for Jupyter notebook servers on Kubernetes, demonstrating that Knative-based serverless deployments achieve approximately 3× faster cold-start times compared to traditional Horizontal Pod Autoscaler (HPA) approaches, while enabling scale-to-zero for dramatic resource savings during idle periods. This work has direct implications for the architecture of GPU-enabled kernel infrastructure in educational settings.

Each research thrust connects to one or more concrete projects: