Gatto

The tension. I started the way a designer starts: wanting mockups. But solo, every week spent in Figma is a week without an answer to the only question that matters at first: does the idea hold?

The choice. AI-centred design: generate, test, throw away. Testing my ideas and the product itself before giving them too much time. Same logic for the system: no design system, a UI kit. This is build mode; a design system would be a waste of time.

What it cost. The UI, precisely. Gatto’s is AI-generated and it shows: decent, not fantastic. That’s the deal I made: until the product thesis is proven, every hour of pixel-perfect is an hour stolen from validation. I save the craft for projects where it pays off; here, the polished UI will come when the product has earned its designer.

The tension. Build a base of genuinely good places, with no editorial team and without reproducing the very bias I’m fighting.

The options. Manual curation (doesn’t scale) · a single API like Google’s (reproduces precisely the popularity bias) · aggregating dozens of editorial sources.

The choice. A source-first pipeline: every place scanned through the SERP, out of which 70+ editorial sources emerged, and a Signal Index based on convergence: counting the distinct providers that recommend a place, not raw mention volume.

What it cost. Running every place through an LLM is expensive. Places under 4.3 on Google are not scanned. I accept the irony: I filter by the very rating I am fighting, and a few gems slip past me, in the name of viability.

The choice. Two self-hosted n8n pipelines rather than a bespoke backend. The first ingests places: the city is meshed into H3 hexagons and swept through Google Places. The second scans mentions: every place is searched in the Google SERP, and every editorial mention that surfaces is classified by GPT-5 (is this place genuinely recommended, or just cited in passing?), then Claude Haiku generates tags and summaries. Two models, two jobs, each on its strength. And above the database, two multi-agent systems: one writes editorial content from our data, the other hunts for new places.

What was genuinely hard. Not the code: the economics. Sorting 70+ sources that are not worth the same. Optimising every scraping request across AI APIs, Apify, Groq. Fetching and keeping Google Places data fresh without burning the budget, because Google is precisely designed to prevent it. And endless iterations on the scanner prompt: the line between “genuinely recommended” and “just cited in passing” lives in a few sentences of instructions.

What it cost. Whole evenings hunting down why nothing was updating, and a discipline kept ever since: distrust silent success as much as loud failure.

The tension. A “96/100” is instantly readable, satisfying, a dopamine hit. It’s the obvious UI. And it’s a trap.

The choice. Internal score only. A number invites gaming, implies false precision, and becomes legally radioactive next to a restaurant’s name. Instead: qualitative signals: tier labels (Must-visit, Hidden Gem), convergence counters (“12 critics keep talking about this place”), rankings. Same discipline in language: “suspicious popularity” rather than “tourist trap”.

What it cost. Harder to communicate, less spectacular at first glance, more design work to make the qualitative credible.

The tension. Building the app was almost the easy part. Getting it discovered is another story: an expert-recommendation app is not a query people type.

What failed. I burned budget on Apple Search Ads before understanding the real problem: nobody searches for what I wanted to sell. People type “best restaurant…”, and that query belongs to Google Maps.

The choice. Going where people actually discover places: TikTok and Instagram. Social content is generated from the database, with hard numbers like “37% of Paris’s best-rated places on Google are cited by no guide at all”: the data promotes the product that holds it.

What it cost. An Apple Ads budget turned into tuition fees, and a lesson engraved: you can’t buy search intent that doesn’t exist.

First numbers (2026). Around 100,000 views on TikTok, 15,000 on Instagram, 400 recurring monthly visitors on the site, 40 App Store installs. Awareness is building; converting views into installs is the open front, and I’d rather write it than dress it up.