An LLM-powered Copilot that provides real-time support for math tutors

An LLM-powered Copilot that provides real-time support for math tutors

An LLM-powered Copilot that provides real-time support for math tutors

Designed an end-to-end AI tool that helps tutors guide students more effectively during live sessions through smart suggestions, step-by-step solutions, and adaptive questioning.

Product Type

End-to-end Design, Chatbot, AI

Duration

Jan 2024 - Sep 2024
(9 month)

Jan 2024 - Sep 2024
(9 month)

Jan 2024 - Sep 2024
(9 month)

Team

Bill Guo (Product Manager)
Zach Levonian (ML Engineer)

Responsibility

Research, Product Strategy,
Prototyping, User Testing

Overview
Overview
Overview

Context

Context

Context

PLUS is a tutoring platform led my CMU and Stanford University

PLUS is a tutoring platform led my CMU and Stanford University

PLUS is a tutoring platform led my CMU and Stanford University

PLUS connects highly trained human tutors with cutting-edge, AI-driven software to boost learning gains for middle school students from historically underserved communities. It has more than 10K hours of tutoring every month, helping thousands of middle school students.

Challenge

Challenge

Challenge

In a race against time, tutors struggle to make math stick and engage students

In a race against time, tutors struggle to make math stick and engage students

In a race against time, tutors struggle to make math stick and engage students

Tutors at PLUS conduct 30-minute sessions with about 5 students, giving them only 6 minutes per student on average. At the same time, they often struggle to explain math concepts clearly and efficiently, or to keep students engaged. In this race against time, tutors must maximize both clarity and engagement to ensure each student grasps the material and feel more motivated.

Solution

Solution

Solution

Empowering tutors with an AI Copilot

Empowering tutors with an AI Copilot

Empowering tutors with an AI Copilot

We developed an LLM-powered co-pilot to assist tutors clearly explain math problems, provide effective encouragement, and ask strategic leading questions, ensuring they make the most of their limited time with each student to enhance engagement and learning outcomes.

Impact

Impact

Impact

300 +

300 +

monthly active users

20%

20%

decrease in time spent explaining math concepts

38%

38%

increase in student engagement

Adjust number of steps with a single click

Adjust number of steps with a single click

Adjust number of steps with a single click

Copilot provides detailed, step-by-step math explanations, with answers that users can easily add or reduce as needed

Expand each step to view teaching prompts

Expand each step to view teaching prompts

Expand each step to view teaching prompts

Users can expand each step to find suggested encouragements and leading questions that help students build confidence and think independently

Provide feedback with ease

Provide feedback with ease

Provide feedback with ease

Users can provide feedback by using presets of categories and options, ensuring they are actionable to the engineers while saving their own time.

Research
Research
Research

Observe and Interview

Observe and Interview

Observe and Interview

12 video analysis and 2 focus group interviews

12 video analysis and 2 focus group interviews

12 video analysis and 2 focus group interviews

  1. Understand session structure

  2. Observe tutor and student behaviors and interaction patterns

  3. Identify challenges and frictions

  1. Understand tutors' pain points - when they occur, their frequency, and severity

  2. Understand how they manage challenges and the support available to tutors

  1. Understand tutors' pain points - when they occur, their frequency, and severity

  2. Understand how they manage challenges and the support available to tutors

User Journey

User Journey

User Journey

Insights

Insights

Insights

In-session support is the most valuable

In-session support is the most valuable

In-session support is the most valuable

The numerous pain points and inadequate existing support highlight a significant opportunity for intervention during the session, where help is most needed.

Key challenges lie in math explanation and engagement

Key challenges lie in math explanation and engagement

Key challenges lie in math explanation and engagement

Tutors struggle most with soft-skill-related challenges, such as maintaining student engagement, guiding students through problem-solving, and offering effective praise.

How might we empower PLUS tutors to make tutoring sessions effective and engaging by providing in-session support that addresses their most critical needs?

How might we empower PLUS tutors to make tutoring sessions effective and engaging by providing in-session support that addresses their most critical needs?

How might we empower PLUS tutors to make tutoring sessions effective and engaging by providing in-session support that addresses their most critical needs?

Ideation and Prioritization

Diverge with AI

Diverge with AI

Diverge with AI

Generating ideas quickly and effortlessly

Generating ideas quickly and effortlessly

Generating ideas quickly and effortlessly

To quickly generate a wide range of creative ideas, I decided to leverage AI-driven brainstorming.

After going through 3 iterations in prompt engineering where I simultaneously evaluate input and output to find the most effective prompt that lead to most reliable ideas, I successfully generated over 200 ideas within just 1 hour.

After going through 3 iterations in prompt engineering where I simultaneously evaluate input and output to find the most effective prompt that lead to most reliable ideas, I successfully generated over 200 ideas within just 1 hour.

Converge with the Team

Converge with the Team

Converge with
the Team

Evaluate technical difficulty with ML engineer

Evaluate technical difficulty with ML engineer

Evaluate technical difficulty with ML engineer

I facilitated a workshop with the head of product and the ML engineer to assess the technical difficulty of each idea. At this point, we didn't use their input as a strict yes-or-no decision maker, but as a reference to guide our design direction.

Validate with end users

Validate with end users

Validate with end users

I also designed a survey for tutors to assess the relevance (validating needs) and helpfulness (validating solutions) of each idea using Likert Scale.  To make tutors better understand and relate to them, we created textual storyboards in a "Problem-Solution-Resolution" format to provide context. Finally, we plotted the average scores of all ideas on a Relevance vs. Helpfulness matrix.

Consolidate findings and make decisions

Consolidate findings and make decisions

Consolidate findings and make decisions

I cross-referenced the earlier assessed technical difficulty of each idea with their relevance and helpfulness to identify the low-hanging fruit—ideas with the highest impact and lowest technical difficulty.

Selected Ideas

Selected Ideas

Selected Ideas

  1. A step-by-step guide to math problems for tutors to provide explanations efficiently

  1. A step-by-step guide to math problems for tutors to provide explanations efficiently

  1. A step-by-step guide to math problems for tutors to provide explanations efficiently

  1. Strategic leading questions for tutors to ask students instead of offering answers directly

  1. Strategic leading questions for tutors to ask students instead of offering answers directly

  1. Strategic leading questions for tutors to ask students instead of offering answers directly

Rapid Prototype
Concept Validation & Prototype

Co-create with Users

Co-design with End Users

Co-create with Users

Co-create with Users

Conversational prototype using AI to find effective and desirable LLM output

Conversational prototype using AI to find effective and desirable LLM output

Conversational prototype using AI to find effective and desirable LLM output

Although the design direction is clear, we were unsure what kind of content output would be most helpful and meet tutors' needs. To develop a tool that's useful for them, we conducted 5 participatory design sessions, inviting tutors to build the co-pilot together.

Although the design direction is clear, we were unsure what kind of content output would be most helpful and meet tutors' needs. To develop a tool that's useful for them, we conducted 5 participatory design sessions, inviting tutors to build the copilot together.

Insights

Insights

Insights

  1. Users want to see encouragements together with step-by-step guide and guiding questions

  1. Users want to see encouragements together with step-by-step guide and guiding questions

  1. Users want to see encouragements together with step-by-step guide and guiding questions

  1. Users prefer a table view over the default bullet view for its structure

  1. Users prefer a table view over the default bullet view for its structure

  1. Users prefer a table view over the default bullet view for its structure

Low-Fi Prototype
Concept Validation & Prototype

Initial Idea

Co-design with End Users

Initial Idea

Initial Idea

Generated output in a table format

Generated output in a table format

Generated output in a table format

The output table includes a step-by-step guide, encouragement, and leading questions.

Although the design direction is clear, we were unsure what kind of content output would be most helpful and meet tutors' needs. To develop a tool that's useful for them, we conducted 5 participatory design sessions, inviting tutors to build the copilot together.

Limitation

Limitation

Limitation

  1. Table can be cluttered with small text, making it hard to read

  1. Table can be cluttered with small text, making it hard to read

  1. Table can be cluttered with small text, making it hard to read

  1. Table creates a poor user experience when the width is restricted

  1. Table creates a poor user experience when the width is restricted

  1. Table creates a poor user experience when the width is restricted

Explorations

Explorations

Explorations

Progressive disclosure is the answer

Progressive disclosure is the answer

Progressive disclosure is the answer

I thought about different ways to progressively disclose for its proven effectiveness in reducing cognitive load, enhancing readability, and improving engagement through an intuitive and smooth experience.

Pros

Pros

Pros

  • Little info shown at a time

  • Easily scannable and digestible

Cons

Cons

Cons

  • Doesn’t Match Users’ Viewing Habits

  • Poor responsiveness when width is restricted

Pros

Pros

Pros

  • Aligning with users needs of viewing steps first

  • Fewer clicks required to see everything

Cons

Cons

Cons

  • Poor responsiveness when width is restricted

Pros

Pros

Pros

  • Progressively disclosure with users' control

  • Optimized for split-screen use with a single-column layout

  • Progressively disclosure with users' control

  • Optimized for split-screen use with a single-column layout

Hi-fi Prototype
Concept Validation & Prototype
Reflection
Concept Validation & Prototype

If I had a chance to redo the project, I would…

If I had a chance to redo the project, I would…

If I had a chance to redo the project, I would…

It was encouraging to see that tutors found many ideas relevant—proof that we were solving the right problems. Some of these didn’t make it into the final version, but that’s a win in itself: we uncovered meaningful needs. I see this as a chance to go back and design even stronger solutions.

If there were no constraints,
I would…

If there were no constraints,
I would…

If there were no constraints,
I would…

Focusing on the most impactful ideas helped us deliver a lean, effective MVP—but it also opened the door to more possibilities. There’s exciting potential to extend AI support beyond the session itself. Several ideas we didn’t pursue were still strong contenders, and I’m excited by the opportunity to grow the tool in ways that keep meeting tutors where they are.

The next step is …

The next step is …

The next step is …

Shipping the MVP was a great milestone, and now there’s a clear path forward. I’d keep building on what’s working—refining based on real-world feedback, collaborating with engineers to improve the model, and watching tutors interact with the tool to uncover small wins. I’m also excited to revisit ideas we parked early on; now that the foundation is in place, there’s room to expand thoughtfully.

Optimized for mobile - still recommend desktop to view images more clearly 🖥️