
Shipped Product

Deejiar is an iOS app for planning trips, with an AI butler and a map of hand-picked places in Tokyo, Singapore and Taipei. I took it from idea to App Store: product direction, design, engineering and go-to-market.
A new class of travelers can afford a great trip but can’t afford to waste one: a long flight, limited time off, no second try.
Travelers a year
~5M
English-speaking, aged 20–40, household income above $100K, visiting Asia's top 10 cities.
Spend per trip
$3,000–5,000
On what they can book, before flights.
Bookable market
≈$15B a year
5M travelers at the low end, $3,000 per trip.
My top-down estimate from tourism-board and industry reports, not yet validated with these travelers.
Travel advisors
Trusted and personal, but priced for the top of the market.
Reviews and rankings
Open to everyone, and 54% of UK travelers worry they're fake or AI-generated.
AI planners
Instant and nearly free, but a guess about a place it has never been.
Icelandair survey of 2,000 UK travelers.
What’s missing is the middle: a day shaped by someone they trust, at a consumer price. In the same survey, 58% now rely on word of mouth from family and friends.
The best journeys aren't planned. They're inherited.
Take a day someone shaped and make it yours. Everything else in the app either starts a day when there’s none to inherit, or carries it through the trip.
Tell Deja what to change. You get your own copy of the day, with every change labeled and one tap to undo.
Every stop opens to photos, what to order and directions. The map shows each day’s stops, in order.
Legs line up by day. One that lands outside your dates gets flagged, and connecting legs join into one trip.
As you pick, the title reads each date back (Oct 20 → Nov 7), so a wrong one shows up before you confirm.
AI does the middle of the work. I own both ends: what it should make, and whether it did.
Deja's body is an equation.
The lattice
y = 1 − 2(i + ½) / n
r = √(1 − y²)
angle = i · π(3 − √5)
Every dot’s position comes from one rule. The golden angle spreads the dots evenly, so no seam or pole gives the sphere away.
Drawn by hand, every size and frame would be its own artwork; here one rule runs live at any size. AI let me work at that level: I described a behavior, read the result and changed the numbers. Where math has no opinion — the eyes, the expressions, a five-second Thinking loop choreographed in Figma — I drew it by hand. The math decides the body. I decide the personality.
Golden angle
Each dot turns 137.5° from the last.
Sphere
300 dots, evenly spread.
Depth
Size and opacity only, no shading.
Eyes
Two ovals, sized as ratios of the radius.
Dot lattice adapted from thinking-orbs (MIT), by Jakub Antalik and Alex Brinza.
Judgment sits at both ends: /plan says what to build, /inspect checks what came out, and /build in between may only stop and ask.
The nine commands are open source: coding-agent-workflow on GitHub
Deja, Deejiar’s in-app butler, turns a line like “4 days in Taipei in November, love food” into a day-by-day Journey. It’s told to build from Deejiar’s own curated places and search the web only when those run thin. So a better model has to be better at that, not just newer.
I measured it as the Deejiar share: the percentage of a Journey’s stops that are real Deejiar places. The bar is 70% in Taipei, where coverage is strongest.
The model proposes · OpenAI
Ask, then draft
Asks where, how long and what you like, then drafts the days and Stops.
Edit by name
Changes a Journey through named edits (add, move, replace a Stop), never by rewriting it.
Search when thin
Uses the web only where Deejiar’s own places run out.
Code decides · Python
Places are checked
A Stop keeps a Deejiar place only if this turn’s facts carried it; the rest are matched by name, then pinned with Google.
Dates and limits
The server sets every date; at most 7 days and 8 Stops a day. A call that breaks a rule is refused, with no Journey and no credit.
Copies, never overwrites
A visitor’s change lands on their own copy; the source Journey is never written.
Every model got the same five requests through one harness: the production planning turn, sent straight to the model at its default effort and scored by the app’s own plan-building code. gpt-5.4 scores the same through the real API route, so the harness is faithful. Tokyo and Taipei pit Deejiar places against the web, Lisbon has no Deejiar coverage, Kyoto gives no dates, and a vague “somewhere in Asia” should end in questions. A run passes when every request ends in questions or a Journey, Taipei reaches 70%, and every Journey is private.
| gpt-5.4default (none) | gpt-6-astradefault (medium) | Sonnet 5.5default (high) | Opus 5.5default (medium) | Haiku 5.5default (medium) | |
|---|---|---|---|---|---|
| Tokyo | 100%7.5 s | 55%23.8 s | 29%32.6 s | 75%38.7 s | 22%19.7 s |
| Taipeibar: 70% | 100%10.9 s | 75%39.6 s | 100%21.3 s | 100%23.0 s | 100%19.8 s |
| Lisbon0% expected | 0%9.5 s | 0%26.2 s | 0%24.2 s | 0%31.8 s | 0%17.9 s |
| Kyoto | 100%5.5 s | 83%26.1 s | 88%19.0 s | 75%34.4 s | 83%11.5 s |
| Vague | 64%9.2 s | 33%28.4 s | 80%23.0 s | 50%22.5 s | 78%20.5 s |
| Result | 5/543 s | 5/5144 s | 5/5120 s | 5/5150 s | 5/589 s |
| Tokyo | Taipeibar: 70% | Lisbon0% expected | Kyoto | Vague | |
|---|---|---|---|---|---|
| gpt-5.4default (none)5/5 · 43 s | 100%7.5 s | 100%10.9 s | 0%9.5 s | 100%5.5 s | 64%9.2 s |
| gpt-6-astradefault (medium)5/5 · 144 s | 55%23.8 s | 75%39.6 s | 0%26.2 s | 83%26.1 s | 33%28.4 s |
| Sonnet 5.5default (high)5/5 · 120 s | 29%32.6 s | 100%21.3 s | 0%24.2 s | 88%19.0 s | 80%23.0 s |
| Opus 5.5default (medium)5/5 · 150 s | 75%38.7 s | 100%23.0 s | 0%31.8 s | 75%34.4 s | 50%22.5 s |
| Haiku 5.5default (medium)5/5 · 89 s | 22%19.7 s | 100%19.8 s | 0%17.9 s | 83%11.5 s | 78%20.5 s |
Share of each Journey's stops that are Deejiar places, one run per cell, with that request's model time beneath.
Speed × token cost
One forced create_journey call
gpt-5.4’s default effort is none: it answers without reasoning. Every other model reasons by default and took 2–3.5× its model time for the same five requests; models that searched the web more also drew fewer stops from Deejiar.
Every model passed 5/5 with no broken tool call, but none matched gpt-5.4, the only one at 100% in Taipei, Tokyo and Kyoto. Production still runs gpt-5.4; a challenger that wins here is confirmed through the real API route before it ships.
What changed is the cost of asking. The harness tests a new model in minutes, without touching production.
Sixteen releases in under five months. Speed only matters if it changes what you do next: in that window Deejiar moved from guides to travelers, from a web wrapper to native iOS, and from two Journey formats to one.

Deejiar is live on the App Store. 40 people have signed up, and a few already plan their trips in it.
Next: the first stranger who pays for Member. A friend has paid before; a stranger paying would show the product made the sale, not the friendship.
