AI-native flows at Escargot
Escargot is a US startup that helps people stay close to their friends and family. Keeping in touch is hard: friends move away, birthdays slip by, and a text gets lost in the feed. Escargot makes it easy to send a real greeting card, a cute reminder that you thought of someone, which they can hold and stick on the fridge. So instead of building a "physical API" to the US Postal Service, our goal was to help people keep their friendships alive.
My responsibilities
Designed and shipped production code with Claude Code, Codex: native iOS in SwiftUI and Xcode, the Next.js website and the NestJS backend. I worked straight in the repository through GitHub, from code review to release, with 82 merged pull requests overall.
Owned whole product areas from 0 to 1 in a team of four: from the first idea and spec to production; web and native iOS app.
Drove decisions with data and research: product hypotheses, user interviews, funnels in Mixpanel and post-launch reports that set the next iteration.
Results
Built AI-native flows with an LLM as the core feature. I prototyped them as working code with AI agents rather than as static Figma screens. 58% of people who got an AI-written card message kept it, and AI gift finder orders averaged 3.5× a typical order.
Shipped native iOS features from Figma to production: the AI gift finder, the AI writing helper, a gift upsell and an adaptive onboarding. All of them were coded in SwiftUI with Claude Code and Codex.
Build the website from a placeholder to a full e-commerce store, designing it and coding most of it myself: design system, tag system, catalogue, checkout, CMS and SEO/GEO. The share of homepage visitors who reach a card grew from 1% to 15%.
LLM Gift Finder
Escargot lets you send a physical gift with the card, but people rarely know what to pick from a catalog. They know the person: she loves cooking, he just moved. The Gift Finder turns that kind of description into a real gift, sent together with the card.
Video
Claude Code, HyperFrames, FFmpeg
How it works
A Gifts card on the Home screen opens a chat. Claude Haiku, connected through OpenRouter, asks about the recipient: who they are, what they love and what the occasion is. Two or three questions are usually enough.
Video
Claude Code, HyperFrames, FFmpeg
Then the search begins:
The LLM turns the answers into search queries.
The backend finds real products in the Shopify catalog, filters them by price, seller and condition, and puts together a shortlist of three gifts.
The LLM writes a short note on why each gift suits the person.
If none of them fit, the user can ask for more options, and the search runs again without repeating earlier picks.

Results:
9.4% of people who saw the Gifts card tapped it.
98% of people who described a friend got suggestions, and half asked for more.
Finder orders averaged 3.5× a typical order, all from first-time gift buyers.
Together with the checkout gift offer redesign, gift orders grew from 0.8% to 5.4% of iOS orders.
LLM Note Helper
Writing the message is the hardest part of sending a card. Blank-card paralysis is one of the main reasons people drop out of the flow and never send the card. I built an AI helper whose messages read as if a friend wrote them.
Video
Claude Code, HyperFrames, FFmpeg
How it works
The flow has two parts. Claude Haiku runs the interview, and Claude Sonnet writes the final message. Both are connected through OpenRouter.
In the first part, the app gathers context about the recipient: shared memories, what they love and more. The user then picks a tone and adds a name. The interview follows a guideline for what makes a good greeting: one true thing about the person, what the message should center on, then a detail in the user's own words. Based on that and the answers so far, the LLM decides which question to ask next.
Video
Claude Code, HyperFrames, FFmpeg
Sounding like a friend
In the second part, Sonnet writes the message from what Haiku gathered. The hard part is making it sound like a friend. I put together a guide to the typical tells of AI-written text and turned it into hard rules in the prompts, so the final message sounds genuine and human.
Results:
31% of people writing a card message opened the helper.
73% of those who started got a message, and 58% of them used it.
Weekly use grew through the first month.
How I worked: designer as product builder
At Escargot I worked as a product builder: one person taking a feature from the first hypothesis to code in production. AI agents were part of every step, from reading analytics to reviewing pull requests.
Step 1. Start from data
Every feature started with a question I could answer myself. I pulled the funnels in Mixpanel, wrote the hypothesis, and after launch came back with a report that set the next iteration.
Step 2. Specs written for agents
I wrote specs that an agent could build from: decisions, edge cases, analytics events and a build order. In Figma I left notes addressed to Claude, so the design file and the code spoke the same language.

Step 3. Prototype in code, then build for real
An AI flow can't be judged from static screens, so I prototyped it in code. The same agents then built the production version in SwiftUI, Next.js and NestJS, while I steered and reviewed every change.


Step 4. Ship like an engineer
Every pull request went through AI review by Codex, Copilot and Claude before merging. One night I ran an agent that checked the whole B2B interface against Figma, fixed the gaps and left a pull request ready for review in the morning.
Results:
82 merged pull requests across 16 projects, on iOS, web, backend and CMS.
Claude Code and Codex were my everyday tools, from spec to code review.
Working this way moved my responsibility from screens to a working product and its numbers.
