Pico

The wager

Pico went from idea to live in São Paulo in one week.

But the wager isn't the shipping speed — it's the format. Pico is a bet that restaurant discovery in São Paulo and Rio works better as a conversation than as a list. It isn't a directory. It's a chat that builds a taste graph as you go: you tell it what you want, what you've liked before, who you're going with, and it answers in two or three places with the specificity that a friend who lives in the neighborhood would.

Google Maps gives you fourteen options ranked by distance. Yelp ranks by stars. iFood ranks by what's commissioned. Pico ranks by what fits you, this evening, given everything you've told it so far — and the more you use it, the better the answer.

The directory format had its run. The next generation of locals will discover restaurants the way they ask their friends: by talking.

How it's built

The headline architecture is conversational discovery on a curated corpus. The piece that's worth talking about is the editorial pipeline that keeps the corpus honest.

Every restaurant Pico recommends has been written about by an agentic pipeline and scored by an LLM-as-judge before it ships. The judge isn't a vague "is this good?" check — it scores on five weighted criteria:

• Specificity (30%) — concrete dishes, ambiance, prices. No generic filler ("a great spot for any occasion").
• Tone (25%) — friend-recommending voice with personality. The kind of phrasing a paulistano would actually use.
• Accuracy (20%) — every claim traceable to input data. No hallucinated dishes, no invented prices.
• Portuguese (15%) — natural Brazilian Portuguese, grammar, cultural fit. Not translated American restaurant copy.
• Structure (10%) — all fields well-formed: tagline, description, vibe tags, best-for occasions, SEO slug.

A restaurant entry that scores below threshold goes back through the pipeline for revision. A restaurant that scores below threshold twice gets flagged for manual review. The pipeline doesn't ship copy it doesn't believe in — which is the only way to build a discovery engine that's trusted instead of skimmed.

The bilingual layer is the other half of the architecture: Portuguese is the primary locale, English is the localized variant. Most of the readers are paulistanos and cariocas. The English exists for the tourists, the expats, and the visiting in-laws.

Underneath: Next.js, Claude Sonnet for the conversation, a Postgres-backed taste graph that updates weights exponentially per signal, and a WhatsApp Cloud API integration that runs the post-visit feedback loop ("did you go? how was it?").

What's working

Since launch on February 14th, 9,200 new users have used Pico to find a restaurant in São Paulo or Rio. That's roughly 100 new users a day, with no paid marketing — entirely organic word of mouth and direct shares.

The corpus is 5,000 places strong — comprehensive enough to feel useful in the neighborhoods that matter, small enough that every entry has been evaluated by an LLM judge before it shipped. Each place has a tagline, a description, a vibe profile, and an SEO-ready slug. None of it is template filler.

The taste graph is doing what it was supposed to do: returning users get faster, more specific recommendations. Someone who told Pico last week that they liked Tan Tan but found Z Deli too crowded for a date will not get told to go to Z Deli for a date. That's the bar.

The post-visit WhatsApp loop is the retention bet. Three days after a recommended visit, Pico checks in. About a third of users respond. Each response improves the taste graph for that user and adjusts the global signal for the restaurant.

What's next

Rio is live. The next two cities under evaluation are Lisbon and Belo Horizonte — both because of demand from existing users, not because of a growth model. The corpus expansion in each new city goes through the same five-criteria judge pipeline; nothing ships before it reads honest.

The transactional layer is the unfinished bet. A smart "Quero ir" CTA, an occasion planner that schedules a visit with friends and settles the Pix afterward, and tighter integration between the chat and the reservation surface. The conversation works; the next step is closing the loop between finding the place and going to the place.

Pico is the cleanest single-product expression of what Frenti does: ship fast, write everything with discipline, let the judge decide what's worth showing, and make every visit better than the last.