Case 01 · B2C · AI · Built · 2026 · 6 min read

Atélier: an AI notes app for iPad

Handwriting-first notes you can search by meaning and summarise in one tap. I took it from a problem I lived as an MBA student to a feature-complete iPad app, solo.

Skip to the decisions

Problem

Students write pages of notes on iPad and cannot find them again. Keyword search fails when you do not remember the exact words you wrote.

What I did

Ran discovery with classmates, bet on search by meaning, cut v1 to two AI features, and designed and built the app in Swift.

Status

Feature-complete and demoed to IIM Lucknow faculty in October 2026. No user results yet: a 20-query search eval and a 10 to 15 student beta come next, with targets set below.

Role: Solo: research, product, design, build (Swift)
Platform: iPad with Apple Pencil
Timeline: July to October 2026
Status: Feature-complete; faculty demo, Oct 2026
Skills: Product sense, AI product judgment, 0-to-1 execution

Context

Why this, why now

At IIM Lucknow most of my cohort takes notes by hand on an iPad. Classes move fast, quizzes come often, and the notes pile up across subjects and weeks.

Two things changed recently. Handwriting recognition on iPad became good enough to trust, and language models made it practical to search a person's own notes by meaning, not just by matching words. That opened a gap: handwritten notes are still treated as pictures you scroll, not knowledge you can ask.

Problem

Pages of notes, never found again

In conversations with classmates before quizzes and case discussions, the same three frustrations came up.

Search fails on memory. People remember the idea from class, not the exact words they wrote, so keyword search comes back empty.
Review is manual. Preparing for a quiz means scrolling notebook by notebook, page by page.
The workaround costs thinking. Switching to typed notes makes search easy but loses the sketches, arrows and structure that come with writing by hand.

Users

Who it is for, and who it is not

The primary persona's moment before a quiz, today and with Atélier. The whole difference is one step: asking by meaning instead of guessing the exact word.

Segment	Job to be done	Call
MBA and grad students (primary)	"Find what the professor said about pricing in week 3, and give me the gist before the quiz."	Primary. Heavy iPad note-takers, high review pressure, easy for me to reach and test with
Working professionals	Recall decisions from meeting notes	Phase 2. Strong need, but notes are spread across many tools
Designers and visual thinkers	Sketch and annotate	Served by the core tools only. They value the canvas more than search

Competition

Where the existing apps stop

Product	What it does well	Where it stops
Apple Notes	Free, built in, searches handwriting	Search matches words, not meaning
GoodNotes	Best-in-class notebooks and handwriting search	Finding still depends on remembering the right word
Notability	Handwriting plus audio recording, searchable	Same keyword limit; review is still scrolling
Typed AI notes (e.g. Notion)	AI search and summaries	You have to type, which loses the handwritten thinking

What I found when I tried these apps in July 2026: handwriting search in all three matched the words on the page, not their meaning. Atélier tests whether search by meaning on handwritten notes is worth building.

Key insight

People remember the idea, not the words.

A student searching for "price discrimination" may have written "charging different customers differently". So search had to work on meaning, not exact keywords. That one insight shaped the product: search by meaning, and summaries grounded only in the user's own notes.

Goals and metrics

What success means

North StarNotes found per weekly active userNorth Star. It moves only when the core promise works: you asked, and you found it.

Searches per active userDo students reach for search at all?
Search success rateShare of searches where the user opens a result
Summaries used per weekIs the one-tap gist part of review?

Guardrail: share of summaries users rate as accurate. Health: week-4 retention.

Hypotheses

The bets, written down before building

H1. Meaning beats keywords. Search by meaning finds the right note in the top 3 for most real student queries. Target: 16 of 20.
H2. Summaries earn a place in review. Students use a summary at least once in a week with a quiz.
H3. Writing feel is the price of entry. If the pen feels worse than their current app, students never stay long enough to try search.

Prioritisation

What made v1, and what did not

Feature	Value to students	Effort	Call
Pen and paper feel (pen, eraser, colour, lasso, ruler)	Table stakes	Medium	v1, first
Search by meaning	High: the core promise	High	v1
Grounded summaries	High: review before quizzes	Medium	v1
Echoes of past notes	Medium: connects today's note to older ones	Low: it reuses search by meaning on the page being written	v1
Chat assistant	Medium	High	Later
Auto-flashcards	Medium	Medium	Later
Sharing and collaboration	Low for the primary user	High	Not planned

v1 has two AI capabilities: search by meaning and grounded summaries. Echoes are not a third; they run the same search on the page you are writing. Two capabilities are also two things I can evaluate honestly.

Solution

What a student sees

Search your handwritingType what you remember, in your own words. Atélier finds the page, even if you wrote it differently.

Echoes of past notesWhile you write, a quiet hint points to related notes from earlier classes.

Pen and paper feelA pen, eraser, colours, lasso and ruler that feel right before any AI shows up.

How a question becomes an answer. Search and summaries are designed to use only the user's own pages. Design is not proof: the eval checks whether summaries stay faithful, with no missing points and no added claims.

Decisions and trade-offs

Three calls, and what each one cost

01Writing experience first

What I chose Chosen

Built a solid pen, eraser, colour, lasso and ruler toolkit before any AI

What I gave up

Weeks of delay before the riskiest bet, search by meaning, could be tested

Why. If writing feels bad, nobody stays long enough to need search. In hindsight, I would have tested search on classmates' existing notes first (see What I learned).

02Grounded summaries

What I chose Chosen

Summaries draw only from the user's notes

What I gave up

"Smarter" summaries that add outside knowledge

Why. Students need the summary to reflect their class, not the internet. The eval checks whether it does.

03Scope

What I chose Chosen

Search and summaries only. No chat assistant, no auto-flashcards in v1

What I gave up

Feature-list appeal

Why. Fewer AI features, each one tested properly.

Status vs targets

Built; the next test decides if it works

Measure	Target	Status
Feature-complete app, solo	Ship v1	Done, October 2026
Faculty demo	Show to IIM Lucknow faculty	Done, October 2026
Right note in top 3 results	16 of 20 real queries	Not yet measured
Summaries rated accurate	8 in 10 or better	Not yet measured
Week-4 retention	4 in 10 beta users	Not yet measured

Targets are not results. This table will show the eval score, including the misses, when it is run.

What I learned

Test the riskiest bet first

I built the writing core before testing search by meaning, the bet the whole product rests on. A cheaper order: run search on classmates' existing notes first, then build the notebook only if search works.
Writing the targets down before the eval keeps the result honest, whatever it turns out to be.

Roadmap

Now, next, later

Now

Run the 20-query eval for search and summaries
Fix whatever the eval exposes first

Beta with 10 to 15 classmates
Track notes found per user and week-4 retention

Later

Working professionals as a second segment
Chat assistant, only if search proves itself