Back to portfolio Case study · CaseAtlas
Case study 01 · 2026 · Solo full-stack build

CaseAtlas. 31,000+ ways to practice strategy.

A free, AI-powered case interview simulator. It writes an original case for the industry, function and difficulty you pick, scores your answers, and tells you what to fix first. I designed and built it solo in 2026, alongside ISB's executive program in product management, and used it to practice what the program teaches: deciding what to build, for whom, and why.

Role
Founder, PM, full-stack builder
Stack
Next.js 16, Supabase, DeepSeek, Vercel
Status
● Live · early access
Live at
01 · The problem

Case prep tools are expensive, fixed, and go stale.

Practicing business cases is how you get into consulting. The tools for it are subscriptions, typically $30 to $60 a month, built around a library of pre-written cases. The cases are fixed, the difficulty is fixed, and the industry and business problem are whatever the case writer chose.

The deeper problem is that a fixed library shrinks as you use it. Once you have seen a case, you have seen it. And most tools stop at the case itself: you read it, you answer it in your head, and nobody tells you whether your thinking was any good.

02 · Who it's for

People in active prep, not casual readers.

I narrowed the audience on purpose: consulting and MBA candidates who are weeks away from interviews and practicing several times a week. They were also the people around me, so I could watch how they prepared.

That last point set the product's direction. The job is not "give me a case". It is "tell me whether I would have passed, and what to work on".

03 · Positioning

Not a case library. Not a chatbot. A practice partner that scores you.

For candidates preparing for case interviews, CaseAtlas generates an original case on demand and scores the structure of your thinking, not just the answer. Positioning statement
AlternativeTypical costWhere it falls short
Coaching$100 to 300 an hourReal feedback, but not on demand and not affordable for daily practice
Case libraries$0 to 200 a yearA finite set of cases and no scoring. Reading, not practicing
General AI chatVariesWill write a case, but with no consistent rubric and no memory of how you are improving
CaseAtlasFree in early accessOriginal cases, a fixed scoring rubric, progress tracked across attempts
04 · The product today

Generator, case, evaluation.

You choose a difficulty and four things about the case: industry, business function, the kind of tension at its heart, and region. That is 13 industries × 9 functions × 15 tension types × 6 regions × 3 difficulty levels, or 31,590 possible combinations, and each one is written fresh.

Case Atlas case generator with difficulty, industry, business function, tension type and region selectors and a Generate a case button.
The generatorPick difficulty, industry, business function, tension type and region. The case is generated on the spot.
Case Atlas loading state titled Building your case, with scenario frame, stakeholder tensions and data exhibits ticked and discussion questions in progress.
Building the caseA live checklist while the case is written: scenario, stakeholder tensions, data exhibits, discussion questions, final polish.
A generated Case Atlas case, Kestrel Dynamics, at Intermediate difficulty with buttons to go to the six questions or download a PDF.
The caseA fresh case with an executive summary, company background, strategic challenge, stakeholder perspectives, exhibits and a set of discussion questions.
Case Atlas evaluation screen showing a score of 7.6 out of 10 with per-question bars for seven questions, six green and one amber.
The evaluationA score out of 10 with a bar for each question, colour-coded by how it went, and up to three attempts per case.
Case Atlas skills radar for a single attempt with six rated skills listed below it.
Skills for one attemptSix skills rated for the attempt: framework, analytical depth, quantitative rigor, stakeholder awareness, specificity and actionability.
Case Atlas dashboard with a score progression line chart over twelve attempts and a skill breakdown radar.
Progress over timeThe dashboard trend line across graded attempts, and a radar comparing the latest attempt with the all-time average.
05 · Decisions that changed the product

What I had, what I changed, and why.

The product today is not the one I first built. These are the changes that mattered, each one a response to something I saw or learned.

● Changed · D1

From a sign-in wall to "try it first"

Originally you had to create an account before you saw a single case. Now a visitor can generate two cases and get full scored feedback with no account. People have to feel the value before they will hand over an email address, and the value here only shows up after the first evaluation.

● Changed · D2

From a single score to a diagnosis

The first evaluation returned one number. A number tells you how you did, not what to do next. It became per-question scores, six rated skills, a "Fix this first" pointer and the option to reattempt. The feedback loop, not the case, is what brings someone back.

● Changed · D3

From 10 frameworks to 23, linked from feedback

I added a small framework panel as a reference. It turned out to change how people read a case: with frameworks in view they map the case against them instead of just reading. So I invested in what was working, grew the library, and made every framework named in feedback open in place.

● Changed · D4

From placeholder samples to real generated cases

The sample cases are the first proof of quality a visitor sees. Hand-written placeholders made a promise the product might not keep, so I replaced them with real output from the generator.

● Changed · D5

From "free on principle" to free with a purpose

I started out treating free as a principle. I now treat it as a strategy: free in early access, capped at 20 cases a month, to get the first 100 people using it. Their feedback shapes the next release, and their behaviour shows me how the product is really used, when, and by whom. Every case costs money to generate, so the cap keeps that affordable while I learn.

● Changed · D6

The AI model, swapped three times

I moved from a small Mistral model to a larger one for reliability, then to DeepSeek. Because every AI call reads its model from one file, each switch was a decision about quality and cost rather than a rebuild.

● Changed · D7

Share links, private by default

Shared cases were open to anyone with the link. A user's answers and scores are personal, so sharing became private by default, as part of a wider security pass before opening the product up.

06 · What I chose not to build

Three reasonable ideas, cut on purpose.

● Cut · A1

Leaderboards and peer comparison

An obvious way to add stickiness. I cut it because it changes the incentive. Chasing a visible rank is not the same as getting better at case interviews, and the two diverge quickly. The product stays in the service of improvement, not performance.

● Cut · A2

A fixed case library

Curating a library would turn CaseAtlas into the kind of product it was built to replace, and put it in a catalog-size race against platforms that have been building theirs for years. The handful of sample cases on the site are a showcase of quality, not the product.

● Cut · A3

A paywall on day one

Gating evaluations or advanced cases was the obvious early move. I held off because I do not yet know what the product is worth to the people using it, and a paywall would have cut off the usage I need to find out.

07 · Making the AI reliable

Building the app was easy. Getting good cases every time was not.

A prompt that produced excellent technology cases would produce thin, generic ones for niche combinations. The output looked structured but had no real analytical tension. Four things fixed it.

● Fix · S1

A strict contract for the output

The model has to return a case in a defined shape, section by section. Output that does not match is rejected, which pushes the model toward specifics, because vague answers tend to fail the check.

● Fix · S2

Automatic retries

Models occasionally return something malformed. The system retries up to two times before showing an error, and the user almost never sees it.

● Fix · S3

Difficulty built into the instructions

Each level has its own instruction set. Beginner cases state the problem clearly. Advanced cases carry ambiguity, conflicting priorities and trade-offs with no clean answer.

● Fix · S4

Speed where it matters, care where it matters

Writing a case is long and needs to be fast, so it runs with the model's extra reasoning switched off. Scoring is short and needs to be right, so it runs with reasoning switched on.

The lesson: the model is one stage in a pipeline, not the whole product. Reliability comes from the structure built around it.

08 · Pricing

Free for now. Priced once I know what it's worth.

CaseAtlas is free in early access: the full product, 20 fresh cases a month, no credit card. Paid plans come later, and early users keep their access when they do.

That order is deliberate. Setting a price now would mean guessing at what candidates will pay. The first 100 users will show me which parts they return to and how often they hit the monthly cap, and that is the evidence a price should be built on.

09 · What I measure

The numbers that will tell me if it's working.

The product is in early access, so I am not going to quote usage figures yet. These are the measures I will watch as the first 100 users arrive, and what each one tells me.

MeasureWhat it tells me
Graded attempts per user, per weekThe main one. A case that is generated but never answered delivered no value
Guests who finish a first evaluationWhether "try it first" gets people to the moment the product makes sense
Guests who create an accountWhether two free cases are enough to earn a sign-up
Reattempt rate, and score change between attemptsWhether the feedback is specific enough to act on
Users who reach the 20-case capWho the heaviest users are, and the first signal of what a paid plan should include
Cost per generated caseWhether free early access stays affordable
10 · What it taught me

Lessons from shipping an AI product.

The job is rarely the feature people name

People said they wanted more cases. What they needed was to know whether their thinking held up. Building for the second one changed the whole product.

Let people reach the value before you ask for anything

Removing the sign-in wall cost me some control and gave visitors a reason to stay. The ask belongs after the proof.

Free is a decision with a job to do

It is easy to call free a principle. It is more useful to say what it is for, who it is for, and what would make you change it.

Scope discipline is how you ship

No leaderboard, no fixed library, no paywall yet. Cutting them is what let one person ship a product that does its one job well.

11 · PM skills demonstrated

A solo product. The whole job.

Problem framing
Found the real job behind the request: not more cases, but knowing whether your thinking would pass.
Segmentation
Built for candidates in active prep, and said no to the casual reader.
Positioning
Placed the product against coaching, case libraries and general AI chat, with one clear difference: scored, structured feedback.
Activation
Replaced the sign-in wall with two free cases so visitors reach the value first.
Prioritization
Invested in what users were already using, and cut leaderboards, a fixed library and an early paywall.
Pricing judgment
Chose free early access with a cap to learn what the product is worth before charging for it.
Metrics
Defined graded attempts as the measure that matters, with supporting measures for each step before it.
AI product thinking
Treated the model as one replaceable stage in a pipeline, with checks, retries and a single place to swap it.
Execution
Designed, built and run solo: product, AI pipeline, accounts, admin tools, security and tests.

Want to try it? It's free.

Live at caseatlas.org. No install, and no sign-up needed for your first two cases.

Next case study

Slack & the K-12 cyber program · 17 to 57 boards

→

Built with Next.js 16, Supabase, DeepSeek, and Vercel. Designed around one idea: good preparation should tell you what to fix, not just hand you another case.