Lumi: an AI reading companion for cross-cultural literature
Lumi is an agentic reading companion for people reading literature in a second language. It brings the missing cultural context onto the page, only when a reader reaches for it, so nobody leaves the story to go hunting through a dozen tabs.
I worked across ideation, research, IA, branding, and system design. In testing, 8 of 8 readers stayed in the story, at 8.6/10 satisfaction.
It started with my camera roll
My workaround for every unknown word (in the pre-AI age) was a screenshot, and the album never stopped growing. It worked, in a way. I could define every word on a page but still could not tell you what the scene meant. The dictionary was never the thing I was missing.
So I stopped guessing, and started wondering how other non-native readers approach culturally rich literature.
Reading in a second language, the hardest part isn't the vocabulary. It's the world behind the words.
Every time a non-native reader stops to look something up, the story goes cold. The context a native reader feels for free, the history, the subtext, the setting, stays invisible. I tried to understand the gap that stands between the visible and the invisible when reading books of non-native culture.
So I asked five other readers.
I looked at what the tools already do, then at what readers actually do when the page stops making sense.
Kindle, Apple Books, Libby, and Kobo, benchmarked against depth-first products like Blinkist, Readwise, and The Pudding.
Readers of culturally rich literature in a translated language.
What I found
Key insight: readers don't want more information. They want just enough of it, sometimes a picture, to crack the invisible cultural code without leaving the page.
Everyone leaves the page to decode
Four of five reached for an external tool to decode what they read, and every one wanted the help to come to them instead.
I type passages into a search engine. It breaks my reading experience; I don't want to switch context.
A popup or footnote, not super detailed but just something, so I don't fall into a Wikipedia rabbit hole.
My efficiency is reduced.
Readers ask for a picture
Readers couldn't picture an unfamiliar world, and several asked, unprompted, for exactly one thing: a picture.
Visuals would be nice, having a picture to look at.
AI-generated visuals would be good, pictures to visualize it a little, like a Tibetan religious activity.
Pictures of food, of cuisines, will help.
Kindle and Apple Books stop at the words
Every direct reading app nails a clean page and then ends at the text. The depth-first products, Blinkist and Readwise, keep their depth outside the story.
- First-gen Chinese American, working long hours in finance, so his 45-minute evening commute is his sacred do-not-disturb reading time.
- Gravitates toward Asian epics and South American literature, because he wants the deep historical subtext and not just the surface plot.
- Got burned once looking up a character's name and hitting a spoiler about their death, so now he skips the confusing parts, even though it bothers him.
I want the real cultural depth, not just the plot. But stopping to Google every historical reference totally kills my flow.
- An avid traveler who uses her reading list to explore the world when she can't afford the plane ticket.
- Loves global literature but hits what she calls cultural aphantasia: she reads a description of a stew, a fabric, a landscape, and cannot picture any of it.
- Treats reading as social. When a cultural norm confuses her she DMs a friend from that culture to ask if it's authentic, rather than settling for a dry definition.
Reading is, in my opinion, the most accessible form of travel. But when I read a word like 'kimchi' or 'sawadikap,' I don't just want the definition, I want to visualize it.
So I set out to build something that
Sarah needs a way into books she would never have thought to search for.
Chen's commute is sacred. Nothing may send him to another app mid-chapter.
Cultural aphantasia is the gap: show the stew, the fabric, the landscape.
Culture explained without giving away what happens next.
Introducing Lumi.
Lumi brings cultural context onto the page, at the exact moment a reader reaches for it.
The MVP narrowed to three pages: a homepage, a library, and the reader, where you read alongside the Lumi agent.
A reading journey built around discovery
A browse-and-explore flow that reads more like a streaming shelf than a bookshelf, where every card leads with context tags and a tight hook.
[ drop video / screens ]
A library that knows where you are
Books organized by where you are in them, not just what you own, with genre tags that reshape the shelf on a tap.
[ drop video / screens ]
In-context help that never breaks the flow
Highlight a phrase, tap once, and Lumi explains it in a chat thread inside the page. Quick-tap tags and follow-up prompts go deeper without ever leaving the story.
[ drop video / screens ]
Visual grounding, zero typing
Lumi reads the passage and surfaces grounding images on its own, so a reader who has never seen the setting sees it without typing a thing.
A library or an e-reader?
How can we give readers the cultural explanations they ask for, while still leaving room for spontaneous discovery as they read?
To answer that question, I first asked
What does book discovery even look like?
Your profile becomes a world map. Finish a book and it unlocks that country, stamping your passport.
- made discovery feel like collecting, built for Sarah
- solved discovery, not comprehension
Hover a country or a region and the map surfaces the books available from there.
- answered Sarah's ask to browse by country of origin
- too heavy to build for this MVP
- a second answer to the question the passport already answered
Highlight a phrase and Lumi explains it in a chat thread that opens inside the page.
- aimed straight at comprehension, at footnote depth
- risks turning reading into studying
Two of these started in Figma Make.
Both discovery ideas came to life in Figma Make, and both turned out bigger than the time I had. So I went straight to in-page explanation. It keeps the reading flow leaning on the agent, and it means that the moment a reader hits something they don't know, the answer is right there on the page instead of three tabs away.
On the other hand
What should the reader control, and what should Lumi decide?
Every control I added was one more thing to manage before the story. So each candidate got one test: does this help the reader stay in the story, or does it turn reading into studying?
- Reads what is on the page
- Prepares the context and the visuals
- Judges when a picture lands better than a paragraph
- Whether to reach at all
- How deep to go
- When to close it and read on
No highlights, no prompts. Help never appears until the reader reaches for it.
A one-tap card that reads like a footnote and is gone in a swipe.
The full analysis panel. Testing read it as study mode, so it never opens on its own.
One rule underneath all three: the reader always reaches first. Lumi is proactive underneath, secondary on the surface, and transparent at the handoff.
Put together, the whole system fits on one map
Two ways in, one reading loop
Sarah walks in the front door; Chen comes back mid-book. Every path ends by returning the reader to the page, so curiosity never costs the reading session.
8 out of 8 stayed in the story.
Usability tested with 8 target readers on Figma prototypes.
- Readers understood the value on sight: staying in a story while getting cultural context on demand.
- The quick footnote-style cards felt natural, like something they already did by hand.
- The deep-analysis panel read as “study mode,” so it stays opt-in rather than automatic.
Lessons, and what's next
I went so deep on the agentic reading experience that I missed a basic feature, the library filter. UX has to hold the full journey, not just the immersive moment.
The same reader wants deep context in one chapter and a light touch in the next. Designing for the shift led to more adaptive patterns than any fixed persona could.
Users hesitated to lean on the AI when they couldn't predict it. Clear hand-off moments mattered more than polish.
I over-invested in micro-interactions some users never noticed. Benefit-driven labels did far more than polish.
What's next
Not everyone reaches for a tool. A few readers reach for a person: a partner who knows the culture, a friend from the region, a crowd-authored subtitle track. Lumi answers with a model, which scales, but it leaves that human-authority instinct unserved. A community or ask-a-reader layer is the honest next step, and the clearest way to grow past what one model can know.
- Surface reading state on the library cards: progress, completion, and downloads.
- Make the AI hierarchy obvious by renaming Ask Lumi to something like Deep Dive.
That's the end of this story. If you're still curious, here are two more.
