Two Lorenzo California sandwiches on metal trays, wrapped in branded paper — Beverly Hills
April 2026

A Mortadella Focaccia Sandwich, and the Product It Became

By Eva Ouyang

After finishing my MBA at USC, I moved away from Los Angeles.

A few weeks into the new city, I found myself craving something very specific, a sandwich from a place in Beverly Hills called Lorenzo California. Not a vague food craving. An exact one. Focaccia bread, folded Mortadella, toasted pistachios that added a crunch and a faint bitterness, ricotta underneath. The best sandwich I'd had in LA.

I opened Google Maps and searched "Italian sandwich." Nothing useful. I tried "Mortadella," then "focaccia sandwich," hoping the search might index a menu or a buried review. Still nothing. I was heading to New York the following week, surely a city that size would have something. However only one result: a chain I'd already tried in LA. Not even close.

Lorenzo California sandwich cross-section — Mortadella, toasted pistachios, ricotta, roasted peppers on focaccia
Lorenzo California, Beverly Hills — Mortadella · Toasted Pistachios · Ricotta · Roasted Peppers on Focaccia.

That left two possibilities. Either Lorenzo California is the best Italian sandwich shop in America, or there are places just as good sitting quietly on some corner somewhere, undiscovered, unindexed, with no way for someone across the country to find them by describing what they actually want.

I started building Nearby.

The problem with maps

This frustration isn't unique to me. Anyone who has moved to a new city and tried to recreate a feeling, not a cuisine, a feeling, has run into this wall. Rating systems, distance filters, cuisine tags: these answer "what is nearby" but not "what will feel like that place I remember."

What actually makes someone want to go somewhere is rarely a star average. It's the text a friend sends: "the light in here at 3pm is golden, you'll love it." It's the travel note that reads: "the moment you walk in, the air changes."

I wanted to build something that could say that.

A long string of failures

I wrote a PRD with ChatGPT, then went to Google AI Studio to prototype.

The technical parts worked fine. The AI could return place recommendations, filter by location, structure the output. But I kept reading the results and feeling like something was off. The words were accurate. They just weren't the kind that make you stand up and go somewhere. They read like a report. I wanted them to read like a letter from someone who actually knew the city.

Every version I tried landed in one of three failure modes:

"There are several highly rated Italian restaurants nearby…" This is a database, not a recommendation.

"Based on your preferences, I suggest the following locations…" This is an assistant, not a friend.

"The area features a variety of café styles suitable for…" This is a travel guide, not an experience.

I knew exactly where the problem was. I couldn't figure out how to fix it.

The Airbnb shift

Then I thought about Airbnb Experiences.

When you travel somewhere new, an Airbnb Experience pairs you with a local, someone who genuinely loves the city and has real taste. They don't hand you a list of attractions. They take you to the corner bakery worth lining up for at 7am, explain what the light does in a certain alley at 3pm. They're observers of the city, not encyclopedias of it.

That was the frame I'd been missing. I wasn't looking for a better recommendation algorithm. I was looking for a voice, the voice of someone who notices things, who cares about the details, who wants to share what's beautiful about a place.

I used Google DeepSearch to research Airbnb Experience's product language and design philosophy, then fed that research to Gemini to rebuild the prompt framework from scratch.

The output was completely different.

What it sounds like now

Here's what Nearby returns for "a place that takes coffee seriously" in San Francisco:

Nearby · San Francisco · coffee master nearby.travelwitheveai.com
"The city awakens with a quiet promise, and for those who understand the language of the bean, a few whispers beckon."
Nearby

The Alchemist's Brew

Russian Hill · Polk Street

Where meticulous preparation meets minimalist elegance, transforming carefully sourced beans into liquid artistry.

"The subtle aroma of perfectly extracted espresso, a whisper of a bloom, in a luminous space."
Nearby

Ocean's Embrace, Coffee's Warmth

Outer Sunset · Lawton Street

A taste of the coast, known for innovative signature drinks and a cozy, spirited atmosphere.

"The invigorating tang of sea air mingling with the sweet, frothy steam of a unique coffee creation."
Nearby

The Heartbeat of the Bean

Mission District · Valencia Street

A pioneer in the city's third-wave scene, offering expertly roasted single-origin and blends with unwavering dedication.

"The vibrant hum of conversation, underscored by the rich, evolving scent of freshly ground beans and a lively energy."

I shared the prototype with a few friends. Their reaction wasn't "oh, useful," it was "I want to go right now." That's the test that matters.

I deployed it properly with Claude Code.

ChatGPT PRD Google AI Studio DeepSearch + Airbnb research Gemini prompt rewrite Claude Code deploy

What I took away

Voice is the product. Most AI tools default to assistant mode: efficient, neutral, complete. Nearby needed something different, restrained enthusiasm, the cadence of someone who genuinely notices things. Changing the tone wasn't a style decision. It changed what the product fundamentally did. Before the rewrite, the reaction was "this is helpful." After, it was "I want to go." That's not a marginal improvement. That's the whole product.

Specific problems produce real products. Nearby didn't start with "I want to build an AI discovery app." It started with one sandwich, one move, one failed search. The more specific the original frustration, the more precisely you can design the solution, and the more other people recognize themselves in it.

As a non-engineer, the leverage is taste. I didn't write a line of backend code. The build used ChatGPT, Gemini, Google DeepSearch, and Claude Code. What I brought was judgment: knowing when the output felt wrong, when it was almost right, when it finally landed. That turns out to be the hardest part, and the part no tool replaces.


AI has made the distance between a feeling and a product much shorter. What it hasn't changed is how long it takes to know what you're looking for.

Nearby is live at nearby.travelwitheveai.com. If you've ever moved somewhere new and tried to find the version of something you left behind, try it.

About this project
Product
Nearby — AI urban discovery
AI model
Gemini 3 (Google)
Built with
Claude Code, Google AI Studio, DeepSearch
Category
Sensory navigation · Spatial AI · LLM products
Builder
Eva Ouyang, PM & AI builder, San Francisco Bay Area
Eva Ouyang is a PM and AI builder in the San Francisco Bay Area.