Kexin Cha

Stanford Data Ocean

by Stanford University

Designed an AI-powered learning platform for health data analysis, from zero to one

My Roles

Product Design (end to end)

Instructional Design

Product Management

Year

2022–2026

Team

Me (sole designer)

Engineer & videographer (1 each)

Stanford instructor (1)

Data & life scientists (rotating)

Industries

EdTech

Health

01 — Overview

Build career-ready skills in health data analysis.

Stanford Data Ocean delivers hands-on biomedical data analysis courses that empower underprivileged learners to understand human health through data.

I designed the platform and its AI Tutor from zero-to-one, and co-designed three Stanford-certified programs with a rotating team of scientists that includes 31 learning modules. 97% of graduates report skills that advanced their careers.

Published in Communications Medicine (Nature Portfolio), 2025 · Second author

Read the paper ↗

$3.2

million

Tuition saved

Tuition saved

97

%

Career growth

Career growth

85

%

Alumni recommended

Alumni recommended

77

%

AI Tutor satisfaction

AI Tutor satisfaction

Startup Founder

Philippines

Build a company

He applied what he learned from the course to build his company, while reflecting on how scarce bioinformatics education and resources remain in the country.

Research Scientist

United States

Search for a job

What began as a strategic move to sharpen her professional edge quickly transformed into a career breakthrough.

Computer Scientist

United States

Mentor next generation

Originally from Nigeria, he plans to mentor upcoming data scientists and explore the cutting-edge intersection of AI and precision medicine.

Scientist

Pakistan

Upskill on the job

Realizing that her bachelor's degree had left a void in her professional toolkit, she took the initiative to master the interdisciplinary skills that her university don't provide.

Startup Founder

Philippines

Build a company

He applied what he learned from the course to build his company, while reflecting on how scarce bioinformatics education and resources remain in the country.

Research Scientist

United States

Search for a job

What began as a strategic move to sharpen her professional edge quickly transformed into a career breakthrough.

Computer Scientist

United States

Mentor next generation

Originally from Nigeria, he plans to mentor upcoming data scientists and explore the cutting-edge intersection of AI and precision medicine.

Scientist

Pakistan

Upskill on the job

Realizing that her bachelor's degree had left a void in her professional toolkit, she took the initiative to master the interdisciplinary skills that her university don't provide.

Startup Founder

Philippines

Build a company

He applied what he learned from the course to build his company, while reflecting on how scarce bioinformatics education and resources remain in the country.

Research Scientist

United States

Search for a job

What began as a strategic move to sharpen her professional edge quickly transformed into a career breakthrough.

Computer Scientist

United States

Mentor next generation

Originally from Nigeria, he plans to mentor upcoming data scientists and explore the cutting-edge intersection of AI and precision medicine.

Scientist

Pakistan

Upskill on the job

Realizing that her bachelor's degree had left a void in her professional toolkit, she took the initiative to master the interdisciplinary skills that her university don't provide.

02 — Problem

Cost, time, and training make health data analysis skills inaccessible.

Unlocking health data insights requires expertise in biology, data science, coding, and ethics- the kind of interdisciplinary training typically reserved for professional degrees.

Yet most people can’t afford it, many schools lack the data and teaching infrastructure, and those making ends meet simply don’t have the time.

Painpoints

73

%

Cannot afford

21

%

Lack of curricula

Few interdisciplinary curricula

6

%

Limited time

Limited time to study

03 — Easy to start, free to join

Jump into hands-on learning with no setup and no cost.

Most beginners quit before analyzing any data because they get stuck installing tools and fixing errors to set up the cloud environment for analysis.

So we removed setup completely and let learners open a ready-to-use workspace instantly. They could focus on the core user flow:

Most beginners quit before analyzing any data because they get stuck installing tools and fixing errors to set up the cloud environment for analysis.

So we removed setup completely and let learners open a ready-to-use workspace instantly. They could focus on the core user flow:

01

Open a Jupyter notebook

02

Load datasets

03

Analyze data

By the end, 98% of learners felt confident in cloud computing, up from 64% before the program.

By the end, 98% of learners felt confident in cloud computing, up from 64% before the program.

Scale metrics

5,000

+

Global learners

Global learners

90

%

Learners study for free

Learners study for free

106

Countries

Countries

Problem 1 · Setup

Zero setup: a ready-to-use workspace opens in one click.

Before: setting up Python and Jupyter locally meant errors like these.

Python and Jupyter setup errors beginners ran into before the cloud workspace

Before: setting up Python and Jupyter locally meant errors like these.

Python and Jupyter setup errors beginners ran into before the cloud workspace

Problem 2 · Cost

Free to join: a scholarship code for learners who can’t pay.

In survey, 73% of learners couldn’t afford the course. Learners earning under $70,000 a year can unlock the full course at no cost.

04 — Good comprehension

Complex made simple with stories & visuals.

I partnered closely with data scientists and life scientists to create 31 beginner-friendly learning modules, removing dense technical language and replacing it with clear analogies and engaging illustrations so learners could grasp complex ideas quickly.

I required instructors to write learning objectives before creating any content. Then I tested every lesson and quiz against those objectives, and sent back anything that didn’t help learners reach them.

These illustrations went into the AI/ML and Bioinformatics modules. Bioinformatics, the topic learners started least confident in, went on to see the biggest confidence gain of any topic.

Comprehension metrics

2×

Learners confident in Bioinformatics, 48% → 96%

31

Modules co-written with scientists

How every module was built

Objectives first, then content, then testing against them.

Learning objectives · example: AI/ML in medical imaging
O1Explain how a model learns from labeled chest X-rays
O2Split images into training and test data
O3Judge a pneumonia classifier with a confusion matrix
Lessons and quiz
Lesson 1: How machines learn from imagesO1
Lesson 2: Training and test dataO2
Lesson 3: Neural networks for X-raysO1
Lesson 4: Reading a confusion matrixO3
Module quizO1O2O3
Machine learning diagram: input examples, relationship, output labels, illustrated by Kexin Cha
Insulin resistant vs insulin sensitive illustration, illustrated by Kexin Cha
Dataset split into training data and test data, illustrated by Kexin Cha
Convolutional neural network: filters and pooling feed input, hidden and output layers, illustrated by Kexin Cha
The anatomy of a biological neuron, illustrated by Kexin Cha
Confusion matrix: true and false negatives and positives, illustrated by Kexin Cha

Illustrations I drew for lessons like these, across the AI/ML and Bioinformatics modules.

Illustrations I drew for lessons like these, across the AI/ML and Bioinformatics modules.

05 — Growing completion

Learning roadmap turns curriculum into momentum.

I designed the course as a clear roadmap. Each module unlocks the next, and milestone badges make progress visible.

After launch, the surveys and analytics I built in from day one showed where learners still stalled. Below is how I found the cause and fixed it. The 30-minute lessons and content changes made with our instructors improved the program’s completion rate.

I designed the course as a clear roadmap. Each module unlocks the next, and milestone badges make progress visible.

After launch, the surveys and analytics I built in from day one showed where learners still stalled. Below is how I found the cause and fixed it. The 30-minute lessons and content changes made with our instructors improved the program’s completion rate.

Graduation rate

50.5

%

Self-paced

Self-paced

85.7

%

Cohort-based

Cohort-based

5-15

%

Industry norm

Industry norm

What the data showed

Programming lost the most learners: about a third of all dormant learners left there.

Where dormant learners leftFirst 3 months after launch · last module opened, in course order
  1. 01Not started17% of dormant learners35
  2. 02Research Ethics10% of dormant learners22
  3. 03Programming32% of dormant learners68
  4. 04Data Analysis4% of dormant learners9
  5. 05Cloud Computing0% of dormant learners1
  6. 06Bioinformatics9% of dormant learners20
  7. 07Survey or certificate8% of dormant learners18
  8. 08Completed journey18% of dormant learners39

What learners told me

The lessons didn’t fit the time they had.

01
Lessons felt too longInactive learners asked us to shorten the content.
02
Learning happened in short gapsThey fit learning into fragmented time during their day.
A learner’s day, illustrated
One long lesson fits none of the gapsSplit into 30-minute parts, it fits

What I changed

30-minute lessons, built in Markdown with no engineering time.

Learner view: lessons split into 30-minute parts

A lesson notebook’s opening as learners see it: objectives, then the module’s lessons (outlined), each about 30 minutes, with “You are here” on the current one

How I built it: Markdown in the Jupyter notebook

The same opening in the notebook’s Markdown and HTML source, with the lesson list outlined

The outlined lesson list is the same in both: what learners see, and the Markdown I wrote for it.

The outlined lesson list is the same in both: what learners see, and the Markdown I wrote for it.

06 — Measure learning

Measure learning from day one, not just completion.

Before the MVP, our instructional design advisor taught me to write learning objectives first and build the evaluation on them. So I designed short quizzes, a final certification exam, and pre- and post-program surveys into the product from day one, to measure skill and confidence.

Before the MVP, our instructional design advisor taught me to write learning objectives first and build the evaluation on them. So I designed short quizzes, a final certification exam, and pre- and post-program surveys into the product from day one, to measure skill and confidence.

Learning outcomes metrics

70

%

Test accuracy

Test accuracy

94

%+

Confident in every topic, after

Confident in every topic, after

97

%

Career growth

Career growth

How I measured

Quizzes, a final exam, and pre- and post-program surveys, built in from day one.

07 — Scalable support

AI Tutor: Scaling student support without scaling cost

With limited budget to hire instructors across time zones, I designed a GPT-4–powered AI Tutor so learners could get instant help anytime. The goal is to reduce drop-off, keeping students supported, and maintaining a high graduation rate without expanding the teaching team.

With limited budget to hire instructors across time zones, I designed a GPT-4–powered AI Tutor so learners could get instant help anytime. The goal is to reduce drop-off, keeping students supported, and maintaining a high graduation rate without expanding the teaching team.

AI Tutor performance metrics

98

%

Accuracy

77

%

Satisfaction

Satisfaction

100

%

Effective guardrail

Effective guardrail

What learners asked

39.5% of questions were programming help, right where learners had dropped off.

What learners asked the AI tutor2,082 questions from 156 learners · share of all questionsProgramming help39.5%

Quiz and exam questions (18%) are one reason the tutor has guardrails. When a question lacks context (9%), it asks for more instead of guessing.

Quiz and exam questions (18%) are one reason the tutor has guardrails. When a question lacks context (9%), it asks for more instead of guessing.

Behind the work

How I worked

No PM, so I ran the roadmap with our stakeholder.

Our stakeholder was a Stanford instructor who brought the funding. We built the roadmap together, and I ran the team from a Notion database of weekly tasks and deadlines.

I co-wrote all 31 modules with our scientists and tested every one myself. With no biomedical data science background, I treated anything that confused me as something that would confuse our learners.

What I got wrong

I scoped the AI tutor for a team we didn’t have.

My user surveys and interviews made the case for an AI tutor, and in early 2023 we won $50K to build it. When graduates asked for more exercise questions, I designed a version where learners generate quizzes from the tutor.

But our only engineer was spread thin from multiple projects, and an AI feature needs constant tuning and UI updates. I hadn’t weighed that cost before pitching it. So I changed course: instructors added exercises directly inside the Jupyter notebooks with no UI change.

What I’d do first next time

Build the shared view of the data before the first decision.

The platform had no analytics dashboard. The drop-off analysis above ran on CSV exports our engineer pulled once a month, which I worked through in Google Sheets. Every optimization decision waited on that handoff, and I accepted that cadence for longer than I should have.

Today I’d make a shared analytics view among the first things the engineer builds, so the data drives the roadmap instead of confirming it.

Next case study

Next case study

applyjobs.club — An AI agent you can see, check and steer

Read the case study →