mahnoor·fatima

MENTOR

The MENTOR interface. A welcome and design challenge screen with a brief and key considerations, a dialogue transcript between student and agents with a design phase progress bar beneath it, and three generated architectural renderings labelled ideation, visualization and materialization.

A multimodal AI teaching system for reflective architectural learning.

MENTOR is an AI-powered educational system designed to support deeper cognitive engagement in architectural education. Rather than functioning as a question-answering assistant, MENTOR guides students through structured reasoning, dialogue, and visual interpretation to encourage reflection, discovery, and independent thinking.

The Story

Developed and evaluated over nine weeks, the system integrates large language models, vision-language models, and agent-based orchestration to scaffold learning across conceptual, spatial, and technical dimensions. MENTOR was benchmarked against a GPT-only baseline, with results demonstrating improved engagement, reduced cognitive offloading, and increased depth of design reasoning.

System Architecture

MENTOR is built as a coordinated multi-agent system, where each agent supports a distinct pedagogical role.

Core Agents

System flow diagram of ten numbered stages, running from the Streamlit interface and state manager through the context reasoning, Socratic, knowledge synthesis and metacognition agents to the LangGraph orchestrator, data module and evaluation module. Each stage is annotated with the Python file that implements it.
What this diagram says

Stages in order: 1 UI (app.py, Streamlit). 2 State manager (state_manager.py). 3 Upload, upload with brief, or brief alone. 4 Context reasoning agent, fed by the vision module (analysis_agent.py, GPT Vision and SAM) and the context module (context_agent.py). 5 Socratic agent (socratic_tutor.py). 6 Knowledge synthesis agent (domain_expert.py, ChromaDB and vector embeddings) with the knowledge manager (knowledge_manager.py). 7 Metacognition agent (cognitive_enhancement.py). 8 Orchestrator module (langgraph_orchestrator.py, LangGraph, LangSmith). 9 Data module (interaction_logger.py, CSV and JSON session export). 10 Evaluation module. Test condition 1 runs the full agent chain, test condition 2 a direct GPT-4 API call, and test condition 3 a set of hard-coded questions.

  • Context Reasoning Agent. Interprets user input and situational context, acting as the entry point for all interactions.
  • Knowledge Synthesis Agent. Connects ideas across domains, synthesizing theory, precedent, and technical knowledge.
  • Socratic Dialogue Agent. Drives questioning, reflection, and conceptual exploration through guided prompts.
  • Cognitive Enhancement Agent. Regulates cognitive load, pacing, and depth of engagement.
Context Agent
Risograph-style illustration in purple and red: a figure in profile with a translucent cube above the head, dissolving into grain at one corner.
Socratic Tutor
Risograph-style illustration: a classical bearded bust in profile, set inside an arched niche against a red and violet gradient.
Domain Expert
Risograph-style illustration: a figure looking down, cupping a glowing orange sphere in both hands, with faint concentric circles drawn behind the head.
Analysis Agent
Risograph-style illustration: a hooded figure in profile beside a vertical strip of spheres shown in successive phases, like a measurement scale.
Cognitive Enhancement
Risograph-style illustration: a figure in profile whose crown is rendered as a dotted woven lattice, framed by a soft-edged rectangle.

Orchestration

The workflows are implemented in LangGraph.

The Context Reasoning Agent routes interactions through single-agent or multi-agent coordination paths, including sequential, parallel, and adaptive flows. A synthesizer stage merges agent outputs into a coherent educational response.

Models.

  • Microsoft Magma for vision-language reasoning
  • GPT-4 Turbo and Claude 3.5 Sonnet for layered interpretation and dialogue

Knowledge Infrastructure

MENTOR’s Domain Expert relies on a hybrid Retrieval-Augmented Generation pipeline.

The pipeline runs in four stages:

  1. Document ingestion. A local PDF repository, text extraction with PyMuPDF, cleaning and preprocessing, then smart chunking with overlaps.
  2. Vector store. Persistent storage in ChromaDB, embedded with sentence transformers.
  3. Search strategies. Semantic search on vector similarity, keyword matching on text, query expansion into enhanced terms, and context analysis of user intent, merged into a single hybrid result.
  4. Knowledge synthesis. Retrieved passages are formatted for readability and citations are tracked, producing a multi-source response.

Educational safeguards.

  • Pedagogical filtering reformats retrieved content for learning contexts
  • Progressive disclosure reduces extraneous cognitive load
  • Citation tracking returns sources in APA, MLA, or Chicago style
  • Results are deduplicated and scored based on authority, relevance, and pedagogical suitability

The design argument underneath all of it is a contrast between two postures an AI tutor can take.

The OracleThe Catalyst
The core transactionStudent asks, AI delivers information. Measures success by answer accuracy. “How quickly can I solve your problem?”Student thinks, AI deepens thinking. Measures success by thinking quality. “How can I make your problem-solving stronger?”
The outcome designFastest path to correct answers. Reduces friction and effort. End goal: task completion.Deliberate cognitive resistance. Maintains productive struggle. End goal: skill development.
The interaction dynamicPositions itself as the knowledge authority. Student becomes the receiver. Creates a dependency loop: question, answer, next question.Positions the student as the thinker. AI becomes the challenger. Creates a growth loop: attempt, reflection, deeper attempt.
What it optimizes forKnowledge transfer, performance optimization.Cognitive development, capability building.

Learning Framework

The system is structured around three learning phases, each with distinct objectives and agent roles.

  • Phase 1: Ideation. Focuses on creativity, conceptual exploration, and questioning. Led by the Socratic Dialogue Agent, with minimal technical constraints.
  • Phase 2: Visualization. Bridges ideas with spatial representation. Balances creative inquiry and technical input using multimodal interaction and analysis support.
  • Phase 3: Materialization. Emphasizes technical knowledge, system integration, and professional practice. Guided primarily by the Domain Expert Agent and real-world constraints.

A weighted completion formula values questioning, conceptual clarity, and response quality, encouraging balanced progression across phases.

Diagram of the three learning phases as columns of coloured boxes. Each column lists its learning objectives, its pedagogical approach, and the order in which the four agents take priority. Two boxes at the foot list transition indicators and transition support mechanisms.
What this diagram says

Ideation aims at program analysis, divergent thinking, vocabulary and framework, taught Socratic-heavy with minimal constraints. Visualization aims at spatial reasoning, representation, integration and communication, with balanced questioning and a multimodal focus. Materialization aims at system integration, construction, performance and practice, led by the Domain Expert against real-world challenges. Transition indicators: conceptual mastery, skill development, challenge readiness, integration ability. Transition support: bridging activities, knowledge consolidation, confidence building, skill assessment.

Interactivity

Task generation. MENTOR generates adaptive challenges aligned with each learning phase:

  • Open-ended exploration in ideation
  • Spatial reasoning and representation in visualization
  • Real-world technical constraints in materialization

Image generation. Visual outputs translate abstract reasoning into diagrams, spatial layouts, and performance visualizations, reinforcing connections between concept and form.

Gamification. Progress metrics, phase completion formulas, and challenge-based learning maintain engagement. Learners advance through levels tied to ideation, visualization, and materialization, reinforcing both creativity and rigor.

Three columns of interface cards showing the same architectural concept development task presented three ways, each with an assignment brief, key considerations, a guidance panel, and a task progress panel.
The same task under the three test conditions: guided Socratic exploration, direct information delivery, and self-directed work.
Six interface panels, each a different challenge format with its own prompt and answer fields.
Six challenge types: role play, perspective wheel, transformation, time travel, mystery investigation, and constraint challenge.

Evaluation and Benchmarking

Cognitive analysis. Student interactions are logged and transformed into interaction graphs. Metrics include:

  • Cognitive offloading prevention
  • Depth of thinking
  • Design move connectivity

Automated fuzzy linkography identifies critical moves and knowledge gaps.

Machine learning integration.

  • Graph Neural Networks detect higher-order cognitive patterns
  • Ensemble classifiers assess proficiency levels
  • Statistical validation using mixed-effects models, ANOVAs, and Bayesian hierarchical analysis

Dashboards visualize learning progress, agent performance, and cognitive metrics in real time.

A grid of twelve screens from the Mentor Cognitive Benchmarking Dashboard, covering key performance metrics, user proficiency, cognitive pattern radar charts, learning progression over time, multi-agent system effectiveness, comparative analysis, personality analysis, linkography, session conclusions, and graph machine learning views.
Twelve views, from key performance metrics through linkography to the graph ML analysis of interaction patterns.

Results

MENTOR demonstrated measurable improvements over a GPT-only baseline.

  • 84.7% cognitive offloading prevention rate
  • 73.9% deeper thinking engagement rate

Students spent more time reasoning through decisions, articulating intent, and reflecting on alternatives rather than accepting generated answers.

MENTOR shows how multimodal, agent-based AI systems can function as cognitive scaffolds rather than content generators. By structuring dialogue, vision, and knowledge retrieval around pedagogical goals, the system supports reflective, discovery-based learning in architectural education.

MENTOR is designed to help students think through design decisions, not around them.

Project developed in collaboration with Seda Soylu and Biel Pitman. You can read the original article on the IAAC blog, or the published paper on Zenodo.