From Prototype to Classroom: An Intelligent Tutoring System for Quantum Education

Synthesis: ITAS (Intelligent Teaching Assistant System) — Elhaimeur & Chrisochoides (2026) describe a multi-agent, knowledge-graph-augmented tutoring system for quantum computing education that bridges the gap between dense mathematical formalism and limited qualified instructors. Building on an earlier knowledge-graph-augmented prototype with two specialized Large Language Models (LLMs) agents (a Teaching Agent and a Lesson Planning Agent), ITAS scales this into a production-grade system with four contributions: a five-module QIS curriculum grounded in an information-first framework, a Spoke-and-Wheel teaching architecture with quantum-specialized agents, production cloud infrastructure, and a conversational analytics layer. Piloted in a real quantum computing course at Old Dominion University, the deployment answered the three questions the prototype left open — whether agent specialization solves the reliability problem, whether the system can run in a real course, and whether the instructor gains actionable intelligence.

Quantum education faces a compounding problem: quantum concepts are counterintuitive, the mathematical formalism is dense, and qualified faculty are scarce outside a small number of well-resourced institutions. ITAS addresses all three by combining a structured Knowledge Graph representation of quantum concepts with specialized AI agents, making intelligent tutoring viable in a technically demanding domain. Its design lessons extend well beyond quantum — they speak to how Agentic AI tutors scale from prototype to production in any discipline with scarce subject expertise.

Key Findings

1. Task-based agent specialization solved the prototype's reliability failure. Where a single overloaded Teaching Agent hallucinated at task boundaries (conflating video timestamps with function names, mixing abstraction levels), the Spoke-and-Wheel architecture kept three specialist agents each within a single scope. Across 75 logged student interactions spanning seven question categories, no task-boundary failures of the prototype type were observed — consistent with Kim et al.'s finding that centralized coordination contains error amplification.
2. The system ran reliably in a real classroom at sub-textbook cost. Migrating to cloud infrastructure on Google Cloud Platform delivered flat sub-4-second response latency at classroom-scale concurrency, with per-student cost estimated below a typical STEM textbook. A missing dependency (pylatexenc) caused a Module 3 error cluster that only surfaced under real deployment, illustrating the gap between a working prototype and production.
3. A conversational analytics agent surfaced findings invisible to any static dashboard. Cross-referencing video seek data, code execution logs, and checkpoint submissions, it detected the "dead zone" — students disengaging at the 42-minute mark because later lecture content had no matching exercises — and the passive-consumer archetype, both without exposing individual student data.
4. Graduate QIS learners engaged the tutor as an intellectual partner, not a remediation tool. Students pushed toward group-theoretic definitions and formal characterizations, a mode of interaction extending beyond the conversational Scaffolding model AutoTutor established and one that existing ITS frameworks do not fully account for.

System Architecture

Knowledge Graph Foundation

• Structure: Concept nodes with prerequisite relationships mapped explicitly across the quantum information science curriculum
• Pedagogical use: The tutor traverses the graph to select the next topic based on learner state, enabling structured progression rather than open-ended dialogue (which can confuse novices in quantum topics)
• Advantage: Explicit prerequisite mapping aids navigation of counterintuitive quantum concepts (superposition, entanglement, measurement) — connecting directly to Knowledge Tracing and Learner Modeling and Adaptive Instruction

The paper deliberately scopes the execution layer — real-time teaching, student interaction, and instructor feedback — while deferring the Knowledge Graph planning layer, including adaptive curriculum sequencing, to future work. Session-based state in a cloud-hosted relational database proved sufficient for execution, and the authors note the planning layer is where the graph will be revisited.

Four Contributions

1. Five-module QIS curriculum grounded in Watrous's information-first framework, structuring the domain into a coherent progression
2. Spoke-and-Wheel teaching architecture with quantum-specialized agents — a deliberate move toward more aggressive agent specialization than the two-agent prototype, in response to task-boundary failures observed under prototype conditions
3. Production cloud infrastructure designed for classroom-scale concurrency at "sub-textbook cost" with regulatory compliance built in
4. Conversational analytics layer for instructors and content developers, surfacing curriculum gaps not otherwise visible

Adaptive Components

Spoke-and-Wheel Teaching Architecture

The Spoke-and-Wheel architecture is the direct response to the prototype's core failure, which the authors attribute to cognitive overload rather than model capability. Three specialist agents analyze each student question in parallel — a Video Agent with a timestamp-indexed concept map, a Guidance Agent calibrated for graduate-level mathematics, and a Code Agent carrying a hand-curated Qiskit error catalog — before a Synthesizer Agent integrates their outputs by a priority hierarchy (code errors first, then conceptual gaps, then video references). The three-agent count emerged empirically: two conflated code and concepts, while four added coordination overhead without reliability gains.

Specialization reduces hallucination in three ways: cognitive load reduction (each agent reasons about one task), task interference elimination (the Video Agent cannot invent function names because it never reasons about code), and synthesis filtering (the Synthesizer discards low-confidence outputs before they reach the student). The reliability principle is that an agent which cannot exceed its designated function cannot hallucinate beyond it — the same principle applied in simplified form to the analytics agent, which narrates pre-aggregated BigQuery summaries and never writes SQL or performs calculations.

QIS Curriculum and Checkpoint Design

The curriculum follows Watrous's information-first approach, treating quantum concepts as generalizations of classical information processing rather than departures from it. Its five modules build through explicit mathematical generalization — Single Systems (Dirac notation, Pauli and Hadamard gates), Multiple Systems (tensor products, entanglement), Quantum Circuits (Bell/GHZ states, no-cloning theorem), Entanglement in Action (teleportation, superdense coding), and Quantum Circuit Cutting (bridging the circuit model to limited-qubit hardware) — anchored in two pedagogical pillars of learning-by-doing and project-driven learning.

A distinctive design choice is the checkpoint philosophy: each checkpoint enforces both correct output and correct implementation approach. This matters specifically in quantum education, where a student can reach correct measurement statistics through an implementation reflecting no understanding of the underlying formalism. Checkpoints carry three separate agent instruction sets — autograding (exact pass/fail criteria, output alone insufficient), guidance (hints with a prohibition on writing complete solutions), and debugging (a per-checkpoint error catalog) — providing differentiated Scaffolding at verification, conceptual, and error-diagnosis levels. Submissions were rejected for correct results reached through hardcoded or concept-bypassing implementations, confirming the dual-output-and-approach requirement functioned as intended.

Cloud Migration and Classroom Deployment

Local execution cannot support a real classroom: no concurrent request handling, no fault tolerance, no session persistence across network boundaries, and no security boundary between student code and the host. The migration to Google Cloud Platform delivered four Cloud Run microservices (Teaching Agent, sandboxed Python Execution, Analytics Ingestion, and Autograding) with auto-scaling and minimum-instance configuration to eliminate cold-start degradation. Reliability was decoupled via fire-and-forget Pub/Sub streaming to BigQuery, so an analytics failure never affects teaching and a teaching failure never corrupts the analytics record.

Privacy and compliance were built in rather than bolted on: student code executes in a container-level sandbox, all interaction data is stored under anonymized identifiers, the analytics agent exposes only aggregate patterns, and the deployment complies with FERPA under Old Dominion's institutional data-governance policies.

The deployment involved five graduate students over a single semester, generating interaction events across video playback, code execution, chat, checkpoint submissions, and session management. Behavioral archetypes emerged that only sustained deployment can surface:

• Self-directed learners relied on video and independent coding with minimal AI interaction, using the tutor mainly for error resolution — supporting autonomous learning and independent struggle.
• Tutor-reliant learners treated the agent as a collaborative study partner, pushing toward formal definitions and group-theoretic connections.
• Late engagers showed non-linear trajectories (one skipping a module entirely), a pattern impossible to detect in a system enforcing sequential completion.
• Passive consumers engaged almost exclusively through video playback with no code execution or chat — a failure mode the reactive teaching system cannot address because these students never signal confusion.

Conversational Instructor Analytics and the Blind Instructor Problem

The student–AI interaction space is intentionally protected as a low-stakes environment for exploration and error, where students can ask questions they would not raise in class without those questions appearing in a gradebook or dashboard. This creates what the authors call the Blind Instructor Problem: the system that most needs to inform the instructor is the one least able to expose its data without compromising the conditions that made the data valuable.

The conversational analytics agent resolves this through aggregate pedagogical intelligence. Its signature finding — the "dead zone" — emerged without an explicit query: a cluster of students stopped watching the Module 2 lecture around the 42-minute mark because all four checkpoints covered single-qubit operations while content from 44 minutes onward covered multi-qubit states, entanglement, and Bell states. Students watched until they had enough background to complete the exercises, then disengaged. The instructor revised the exercise set mid-semester, a change motivated entirely by the analytics agent's output. The agent also distinguished failure types — a variable-name error read as a comprehension issue (calling for a clarification note) versus a genuine QiskitError reflecting a conceptual gap (calling for a lecture revision).

Design Principles for AI-Enhanced Flipped Classrooms

Four design principles emerge from the deployment:

1. Exercise coverage must match content coverage. Students treat exercises as the operational definition of required material; any quantum concept without a corresponding implementation exercise risks being skipped regardless of importance.
2. Checkpoints should enforce implementation approach alongside correctness. In quantum computing, multiple implementations can produce identical measurement statistics, so requiring specific construction patterns prevents hardcoding of correct outputs.
3. AI tutors for graduate QIS must match the audience's mathematical sophistication. Graduate students did not ask for simplified intuitive explanations; they asked for group-theoretic definitions and formal characterizations. Prompts calibrated to undergraduate audiences underserve this population.
4. AI-enhanced flipped classrooms require more robust monitoring, not less. Students who would signal confusion in an in-person session may consume content passively in a self-paced environment, making the analytics layer a pedagogical necessity rather than a convenience.

What this means for practice

• Instructors. Map exercises to lecture content before the term starts: ITAS's analytics agent found a cluster of students leaving the Module 2 lecture around the 42-minute mark because all four checkpoints covered single-qubit operations while content from 44 minutes onward covered multi-qubit states and entanglement.
• Instructors. Require checkpoints to verify implementation approach as well as correct output, since a student can reach correct measurement statistics through a hardcoded or formalism-bypassing implementation.
• Instructional designers. Calibrate tutor prompts to the actual audience — the Guidance Agent was instructed that "these are graduate students. Engage with the mathematics; do not oversimplify" — because graduate QIS learners asked for group-theoretic definitions rather than intuitive simplifications.
• EdTech designers. Specialize agents by task rather than scaling one general agent: the Spoke-and-Wheel design runs three specialists (video, guidance, code) behind a synthesizer because two agents conflated code with concepts while four added coordination overhead without reliability gains.
• Instructors. Instrument self-paced courses for aggregate monitoring and act on it mid-semester: the analytics agent surfaced the dead zone and a passive-consumer pattern from video seek, code execution, and checkpoint data without exposing individual students, and the instructor revised the Module 2 exercise set as a result.

Limitations

• There is no control group, so outcomes cannot be attributed causally to ITAS rather than to the curriculum or to the instructor — who designed both, introducing a design and interpretation bias.
• Validity follows qualitative standards here: it rests on the information richness of the cases and the triangulation of evidence sources, not on sample size.
• The deployment covered one course at one institution, so generalizability is not established; the current version shows that the execution layer works, not yet that it teaches better than the alternative.
• A controlled, blinded comparison against the prototype on the same input set, and AI-tutored versus human-TA comparisons with validated instruments, remain future work.

Citation

Elhaimeur, I., & Chrisochoides, N. (2026). From Prototype to Classroom: An Intelligent Tutoring System for Quantum Education.