Miroslav Suzara

Researcher and engineer building AI systems that help people learn, collaborate, and create.

PhD candidate in Learning Sciences and Technology Design · MS in Computer Science · Stanford University

MSc in Learning & Technology · University of Oxford

email · linkedin · scholar · github

I care about whether these systems make people more capable over time, not just more productive in the moment. At NVIDIA, I built and presented demos across NVIDIA’s AI ecosystem for C-suite and enterprise audiences. At Patagonia, I am working directly with senior leadership and operating teams to build and evaluate agentic AI workflows around real operational problems. At Stanford, my PhD work explores how people learn, collaborate, and create with AI.

Projects

TeacherSim

An experimental teaching simulator exploring how educators might practice with AI students. Designed as a provocation for deliberate practice, each session makes pedagogical moves visible for reflection and iteration.

teacher rehearsal simulated students deliberate practice

Explore what practice teaching with AI students could look like

TeacherSim is an exploratory prototype for low-stakes teacher rehearsal. It pairs you with AI students configured with distinct personalities, confusion patterns, and interaction styles so you can probe the design space around practice, feedback, and classroom-like conversations.

Full teaching session UI

Whiteboard, student profile, inferred confusion signals, voice controls, and behavior settings are all visible in one view. The student panel exposes the system's working model of comprehension and personality so you can explore how that kind of instrumentation might shape teaching decisions.

Seven research-inspired personas

Each student persona, from Confused Beginner to Overconfident Misconceiver to ESL Learner, draws on patterns described in education research. They interrupt, go silent, or push back in ways meant to evoke plausible classroom dynamics.

Exploratory student profile

Each session maintains an exploratory model of the student's understanding, rapport, and behavioral state. The sidebar surfaces those inferred shifts as a design probe rather than ground truth, helping test how useful this kind of feedback actually is.

Voice-first, interruptible design

Continuous voice mode auto-sends after a natural pause. Student voice responses are spoken aloud. The interaction aims for something closer to a classroom conversation than a turn-based chatbot exchange.

Configurable sessions

Choose your student persona, pick a CS topic (DFS vs BFS, sorting, recursion, loops), and select a session type (Teaching Section or Office Hours). Each combination creates a different practice scenario.

Stack

FastAPI JavaScript Local LLMs LoRA adapters LM Studio Web Speech API tldraw

ChatCollab

A multi-agent framework where human and AI teammates collaborate as peers inside Slack. Built to study how agent design shapes group dynamics, participation equity, and decision quality. Each configuration generates rich behavioral simulation data on how different team compositions communicate, coordinate, and make decisions.

human-AI teaming team communication group dynamics

paper

Simulate different AI team compositions in Slack and compare how collaboration changes

ChatCollab drops AI agents into Slack as peer collaborators. Configure team compositions, run simulations, and analyze how different agent designs affect participation equity, communication patterns, and decision quality across hundreds of conversations.

Multi-agent Slack workspace

AI agents join Slack as named team members (Product Manager, CEO, Developer), each with defined roles and autonomous decision-making. They collaborate in real channels alongside humans, producing naturalistic conversation data.

Agent interaction sequences

Speech bubbles show real agent dialogue with natural handoffs, from clarification questions to code review to PRD generation. Different team structures produce measurably different collaboration patterns.

Behavioral analysis across conditions

Interaction patterns mapped across experimental conditions, color-coded by message category (orientation, suggestion, opinion, solidarity). The visualization shows how team-level and agent-level guidance shape group dynamics.

Stack

Streamlit FastAPI SQLite SQLAlchemy Slack API OpenAI API Docker

TeacherCollab

A multi-agent simulation where novice teachers develop responsive teaching practices with AI colleagues. Teachers first explore student work one-on-one with each AI persona, then join a roundtable where all three AI teachers and the human co-analyze student thinking together.

teacher professional development student thinking responsive teaching

Practice noticing student thinking with AI teacher colleagues who challenge your reasoning

Responsive teaching (noticing, interpreting, and responding to student mathematical thinking) is hard to practice alone. TeacherCollab pairs you with three AI-simulated colleagues focused on student interaction, problem-solving strategies, and mathematical understanding. First you explore each lens one-on-one, then join a roundtable where everyone co-analyzes student work.

Collaborative analysis workspace

The main interface shows student work alongside a collaborative analysis panel. AI teacher colleagues offer multi-perspective analysis in real time, prompting you to notice mathematical strategies, student reasoning patterns, and pedagogical decision points.

Configurable agent personas

Each AI colleague has a distinct focus, initial prompt, and custom instructions. Teachers can examine and adjust how each agent approaches student work, making the simulation transparent and adaptable to different learning goals.

Real student artifacts

Teachers analyze authentic student work from handwritten math, diagrams, and problem-solving attempts drawn from real-world classrooms. Multiple learning sequences can be configured through administrator settings, allowing teacher training programs to modify scenarios and bring in their own student work.

Stack

Next.js React TypeScript Tailwind CSS Prisma Postgres OpenAI API ElevenLabs

Bruno

An AI-enabled collaborative learning platform where an AI agent facilitates real-time group problem-solving. Bruno joins remote sessions via video, whiteboard, and chat, offering live or reactive feedback grounded in collaboration frameworks to support equitable participation and productive group dynamics.

group problem solving AI facilitation real-time collaboration

An AI agent that joins group problem-solving to support equitable, productive collaboration

Bruno is an AI-enabled platform for real-time remote collaboration. Participants work together via video conferencing and a shared whiteboard while an AI agent monitors conversational dynamics through live transcription, offering feedback to promote shared understanding, balanced participation, and effective team regulation.

Real-time collaborative workspace

Video, audio, and chat in a single interface alongside a shared whiteboard. Participants collaborate through multiple channels while the AI agent monitors diarized discourse and whiteboard activity through live transcription, offering contextual facilitation grounded in collaboration frameworks.

Video conferencing with an AI agent

Bruno AI joins the call as a participant alongside human collaborators. Video, audio, chat, reactions, and screen sharing give groups multiple channels for communication while Bruno monitors and facilitates.

Collaborative whiteboard

The interactive whiteboard supports visual communication and joint meaning-making. Participants externally represent ideas, draw diagrams, and annotate solutions together while the AI agent observes and responds to the collaborative activity.

Stack

React TypeScript tldraw Python FastAPI Daily OpenAI API ElevenLabs

Spellburst

Node-based creative coding with LLMs, built to preserve human agency over the creative process while enabling rapid exploration at multiple levels of abstraction.

creative coding prompt-to-code visual exploration

paper website

Explore creative code at every level of abstraction without losing control

Spellburst is a node-based interface for creative coding with LLMs. Chain prompts, fork explorations, and refine outputs visually, keeping human agency at the center of the creative process while letting AI handle the boilerplate.

Large creative jumps with natural language

Describe what you want in plain language and the system generates p5.js code to match. "Turn the canvas into a vibrant stained glass window" produces working generative art instantly, letting you make big creative leaps without writing code.

Parameterization with semantic sliders

The system auto-generates interactive sliders mapped to semantic concepts from your prompt. "Irregular", "semi-transparent", "colorful" each control a code parameter, letting you fine-tune the output without touching code.

Granular control with code editing

When you need precision, drop into the code editor and modify p5.js directly. The live preview updates as you type. You can seamlessly switch between natural language, sliders, and code at any point.

Stack

React React Flow Jotai d3.js Node.js Express OpenAI API p5.js

Selected Publications

Spellburst: A Node-Based Interface for Exploratory Creative Coding with Natural Language Prompts

Angert, T., Suzara, M., Han, J., Pondoc, C., et al.

UIST 2023

link

The GPT Surprise: Offering Large Language Model Chat in a Massive Coding Class Reduced Engagement but Increased Adopters' Exam Performances

Nie, A., Chandak, Y., Suzara, M., Malik, A., et al.

L@S 2025

link

PearProgram: A More Fruitful Approach to Pair Programming

Bigman, M., Roy, E., Garcia, J., Suzara, M., Wang, K., et al.

SIGCSE 2021

link

ChatCollab: Exploring Collaboration Between Humans and AI Agents in Software Teams

Klieger, B., Charitsis, C., Suzara, M., Wang, S., et al.

arXiv 2024

link