JOLI Collective —
an editorial travel concierge.
A travel concierge that reads like a well-travelled friend's note, not a recommendation feed. Built as my own venture, JOLI is a stress-test of what design-led AI looks like when craft sits next to capability.
Context
Algorithmic travel recommendations have a flatness problem. Everything is good, nothing is interesting, and the things that get surfaced look the same across every platform. JOLI started from the opposite premise: a small set of opinionated recommendations from someone whose taste you trust, in places where taste matters more than ranking. I built it to find out what an editorial AI product feels like when it's designed end to end - voice, surfaces, system, and model layer - by the same hand.
Approach
The brand came first. Monument Grotesk, an editorial reference register that sits closer to Cereal and Monocle than to a typical AI product. Present-tense, opinionated, written in the voice of someone who has actually been there. The product was designed to extend that voice - not to bolt a chat interface onto a recommendation engine, but to feel like a continuous editorial surface that happens to be alive.
The architecture is a multi-stage AI pipeline with three section-scoped copilots — Stay, Eat, Do — each with its own thread, context, and tools that mutate state. The copilots don't share a brain; they share a brief. That separation is what lets each one stay opinionated rather than collapsing into a generic assistant.
What I built
In collaboration with designer Natalia Witwicka. Wordmark, editorial typography system on Monument Grotesk, voice guidelines, photographic register. Reads as taste, not as a tech product.
Multi-stage AI pipeline. Three section-scoped copilots (Stay, Eat, Do), each with isolated thread, context, and tools that mutate state. Designed surfaces, not chat-as-default.
Next.js 14 App Router on Vercel. Supabase with pgvector and PostGIS for geo-aware retrieval. RLS across 22 tables, rate limiting, Stripe webhooks. Claude Sonnet at the model layer with OpenAI embeddings.
Separate codebases for the marketing site and a Remotion-based video pipeline, sharing a TypeScript-native design system. Brand, product, and the surfaces around it, shipped from one stack.
Outcome
JOLI is my own answer to a question worth asking out loud: what does an AI product feel like when it's designed end to end by someone who cares about both halves?
The work continues - new sections, new surfaces, new model layers - and what's shipped already is the proof of approach: idea to sketch to shipped.
Credits
Natalia Witwicka — wordmark, identity system, and editorial typography in collaboration with Flykite.
Flykite Studio.
Meet Milo and other experiments — small builds, shipped to see what sticks.
See more work