I started MonkeyEatingMango to learn AI. A few weeks in, it’s become an obsession — turns out itinerary generation is a surprisingly beautiful problem.
I’ve built an MVP, submitted the sitemap to Google, and made a few comments on Reddit. That alone brings a few users a day, from places I’ve never been. Last week: 96 users from 25 countries, including Brunei, Austria, Hungary, and Bahrain.
Here are the reasons why I think it’s a great idea and my counters to obvious criticisms
What I love about this problem
It’s the right shape for AI right now. Itinerary generation sits in the sweet spot where current AI is genuinely useful — natural language input, structured output, retrieval over factual content, multi-constraint optimization across time, budget, energy levels, group dynamics, geography. Each of those is a problem the field has made serious progress on in the last 18 months. None existed in a usable form three years ago.
Building this product was impossible a year ago. It’s reasonable now. That’s the moment I want to be building in.
It’s also simple enough to ship a useful v1 in weeks, and hard enough in non-obvious ways to stay technically interesting for months. The first version writes itself. The good version takes real engineering — retrieval quality, eval harnesses, fine-tuning on preference data, agentic feasibility checks. I’m building toward that, and I’ll write about each piece as I work on it.
I love the problem and I understand what good looks like. I can’t say that about a lot of other problems. Taste matters, and travel is a domain where mine is reliably calibrated.
It’s bounded and unbounded at the same time. There are maybe 5,000 places in the world people actually want to visit, and all 7 billion humans want to go to them. Long tail exists, but the Pareto principle does real work here. My hunch is that 20% data covers 95% of use cases.
Users seek out the solution. Just check Google search volumes. SEO works more like engineering than B2B sales — you build it well, it compounds. Travel is also visual, so I’ll need video at some point. AI-driven video generation paired with grounded itinerary data should work well for that.
Travel is lucrative. If I build something useful and interesting, it will be financially rewarding. Monetization is not a hard problem in this space, and the cost of running this app doesn’t scale linearly, despite AI integrations. It actually will drop significantly after a point.
What concerns me (and how I counter them)
“Why build when there are so many already?” Wrong framing. Many entrants is the signal — demand exists, no clear winner.
“There’s no moat.” I don’t need one. Execution, customer focus, and creativity are the only moats. And in AI products specifically, the real moat is the system around the model — retrieval pipeline, eval harness, data flywheel, the specific failure modes you’ve measured and fixed. None of that is the model itself, and none of it is downloadable.
“It’s a solved problem.” It isn’t. The impedance between what the AI sees and what exists in the real world can shatter any travel plan.
Case in point: a Spanish football manager lost his job at a Russian club allegedly because he used AI to plan a long away trip, and the resulting itinerary had players staying awake for 28 hours. The plan got shelved before the trip, but the story stuck with me. That’s what AI itinerary generation looks like when it collides with reality.
That collision is the whole problem. The model knows facts that aren’t true anymore. It hallucinates restaurants that closed. It recommends a 4-mile walk between stops it doesn’t realize are uphill. It books your morning at a museum that’s closed on Tuesdays in November. Solving this isn’t about a smarter model — it’s about better retrieval, real-world feasibility checks, and evaluation infrastructure that catches these failures before users see them.
That’s what makes this hard. It’s also what makes it interesting.
Right now, MonkeyEatingMango is a project, not a startup. I’m totally absorbed by it. I’ll see what the signals are and where they lead.
Suggestions welcome!