I11 · Intermediate Capstone: An End-to-End AI Project

Design Thinking and Problem Framing

Design thinking is a human-centred approach to solving problems in five overlapping stages: empathise with the people affected, define the problem clearly, ideate many possible solutions, prototype quickly and test with real users. Problem framing for an AI project turns a broad concern into a precise statement: who the users are, what decision the AI will support, what it will predict, what data exists, how success will be measured and what could go wrong. A well-framed problem often shows that a simple rule or a non-AI solution is enough.

12 min Intermediate · Sublevel Advanced Jump to the activity

What you need: The earlier topics of this module.

Definition

Most failed AI projects do not fail because the model is weak. They fail because they solved the wrong problem, for the wrong people, with a goal nobody could measure. A capstone project starts with framing the problem well.

Design thinking is a human-centred approach to solving problems in five overlapping stages: empathise with the people affected, define the problem clearly, ideate many possible solutions, prototype quickly and test with real users. Problem framing for an AI project turns a broad concern into a precise statement: who the users are, what decision the AI will support, what it will predict, what data exists, how success will be measured and what could go wrong. A well-framed problem often shows that a simple rule or a non-AI solution is enough.

In simple words

A doctor first asks questions and examines the patient before prescribing medicine. Design thinking is the same: understand the people and their real problem before choosing any technology.

Overview

  • Design thinking is a human-centred way to solve problems by understanding users first.
  • The common five stages are empathise, define, ideate, prototype and test (Stanford d.school).
  • Problem framing turns a vague idea into a clear, measurable AI task.
  • A good frame states the user, the decision, the prediction target, the data and the success metric.
  • Many problems do not need AI; a simple rule or better process may be enough.
  • The stages are a loop: testing often sends you back to redefine the problem.

Why it matters

Framing decides everything that follows: what data you collect, which model you build and how you judge it. Time spent here saves weeks of building the wrong thing, and it is what examiners, employers and clients look for first.

Key terms

Design thinking
A cycle of empathise, define, ideate, prototype and test, centred on real users.
Problem statement
A precise description of whose problem is solved, what decision is supported and how success is measured.
"How might we" question
A framing that opens up solutions, such as "How might we warn families a day before air turns hazardous?"
Stakeholder
Anyone who uses, is affected by or is responsible for the system.
Success metric
A measurable target linked to the users' goal, not just model accuracy.
Minimum viable product (MVP)
The simplest version that real users can try and give feedback on.

Types

Design thinking has several types of stage and framing tool. Each produces something specific.

1. Empathise: user research

Learning about the people who have the problem.

  • How it works: Interview users, observe their work, and record pains, goals and context in notes or empathy maps.
  • Example: Talking to five shopkeepers to learn why they run out of popular items.
  • When to use it: Use it at the start of every project so you solve a real problem, not an imagined one.

2. Define: problem statement

Writing one clear sentence that states who has the problem, what it is and why it matters.

  • How it works: Use the form "[User] needs a way to [goal] because [insight]" and turn it into a "How might we…?" question.
  • Example: "Small shopkeepers need to know what to restock each week because running out loses customers."
  • When to use it: Use it to keep the team focused and to check later whether the solution fits.

3. Ideate and choose the AI task

Generating many possible solutions and selecting the right type of ML task.

  • How it works: Brainstorm without judging, then map the best idea to a task: classification, regression, clustering, recommendation or generation.
  • Example: Choosing weekly sales regression for restocking rather than a chatbot.
  • When to use it: Use it to pick a solution that is useful, feasible with your data and simple enough to finish.

4. Prototype

Building a quick, rough version to try the idea.

  • How it works: Make paper sketches, spreadsheets, a notebook with a simple model or a basic web demo.
  • Example: A spreadsheet that averages the last four weeks' sales to suggest restock quantities.
  • When to use it: Use it to learn cheaply before investing weeks in a full model.

5. Test and success metrics

Trying the prototype with users and measuring whether it helps.

  • How it works: Define a model metric (such as MAE) and a real-world metric (such as fewer stock-outs), collect feedback and iterate.
  • Example: Shopkeepers reporting 30% fewer empty shelves after a month.
  • When to use it: Use both kinds of metric: a model can score well yet fail to help users.

Explained in detail

Our running project

Throughout this capstone we follow one project. Many Indian cities suffer from high PM2.5 pollution, tiny particles that damage lungs and hearts, especially in winter. India's National Air Quality Index, launched in 2014, classifies the 24-hour average of PM2.5 as Good (0 to 30 micrograms per cubic metre), Satisfactory (31 to 60), Moderate (61 to 90), Poor (91 to 120), Very Poor (121 to 250) and Severe (above 250). Today most people learn the air is bad only when it already is. Our idea: predict tomorrow's PM2.5 for a city so that schools, hospitals and families with asthma patients can prepare a day early.

The five stages applied

StageWhat we do in the air-quality project
EmpathiseInterview a school principal, a doctor, a parent of a child with asthma and a municipal officer about what they need to know and when
Define"Principals need a reliable alert by 6 p.m. when tomorrow's PM2.5 is likely to be Poor or worse, so they can move outdoor activities inside."
IdeateOptions: a forecasting model, a simple rule based on today's value, SMS alerts, a dashboard, a WhatsApp bot
PrototypeA spreadsheet of past data, a baseline model and a mock SMS message
TestShow the mock alerts to five principals, check whether they would act on them, and refine

From vague idea to measurable goal

"Use AI to fix pollution" is not a project; it is a wish. A good AI problem statement names the user, the decision, the prediction, the timing and the measure of success. For our project the prediction target is tomorrow's 24-hour average PM2.5; the decision is whether to send an alert; and success is measured both technically, as the mean absolute error of the forecast and the recall of Poor-or-worse days, and practically, as whether schools actually change plans. We also write down what failure looks like: missing a dangerous day is worse than an unnecessary alert, so recall of bad days matters more than precision. Finally we ask whether AI is needed at all: if "tomorrow will be like today" were already accurate enough, a simple rule would do, and the next pages test exactly that.

How it works

  1. Talk to at least three kinds of people affected by the problem.
  2. Write a one-sentence problem statement naming user, decision, prediction and timing.
  3. Define technical and real-world success metrics, and which errors are worse.
  4. List possible solutions, including non-AI ones, and pick the simplest promising one.
  5. Build a quick prototype and test it with real users before polishing.

Try it yourself

The design thinking cycle

Put the stages in order.

  1. 1Empathise: meet and listen to users
  2. 2Prototype: build something quick and simple
  3. 3Define: write a clear problem statement
  4. 4Test: try it with users and learn
  5. 5Ideate: list many possible solutions

Well framed or vague?

Classify each problem statement.

  • "Make a cool chatbot"

  • "Alert school principals by 6 p.m. when tomorrow's PM2.5 is likely to be Poor or worse"

  • "Use AI to improve education"

  • "Flag complaints about water supply for the ward office within an hour"

Common mistakes

  • Choosing the technology before understanding the users.
  • Measuring only model accuracy instead of the users' real goal.
  • Skipping the question of whether a simple rule would work.

Check yourself

Question 1

1. What is the first stage of design thinking?

Question 2

2. A good AI problem statement names…

Question 3

3. In the air-quality project, why does recall of Poor-or-worse days matter most?

Summary

  • Design thinking cycles through empathise, define, ideate, prototype and test.
  • A framed AI problem names users, decision, prediction, timing and measurable success.
  • Our capstone forecasts tomorrow's PM2.5 to alert schools, valuing recall of dangerous days.

Finished reading?

Sources: Stanford d.school: design thinking resources · Central Pollution Control Board: National Air Quality Index