AI Travel Concierge for a Consumer Travel Startup

The challenge

Planning a good trip still costs 8–20 hours of open tabs, screenshots, group chats and a spreadsheet nobody maintains. The founder's insight was that this isn't a content problem — the internet already has every recommendation you need — it's a consolidation problem: reviews sit on one site, hours on another, the map in a third, a friend's tip buried in a DM. So the brief wasn't 'a chatbot that writes itineraries.' Generic LLM output reads plausible and falls apart on contact with reality — closed on Mondays, ninety minutes across town, wrong season. The system had to produce a plan that was specific, geographically coherent, editable by a group in real time, and grounded in real place data — fast enough that a daydreaming user doesn't bounce — and it had to earn the right to charge, because a free AI itinerary is a commodity.

The approach

• Treat the LLM as a planner, not an oracle — the model sequences and narrates, while real place data (hours, geo, ratings, categories) is injected from authoritative sources, never invented
• Optimize for the first 60 seconds — stream the itinerary as it generates so the user sees structure forming instead of a spinner; perceived speed is the retention lever at the top of the funnel
• Make the artifact collaborative from day one — an itinerary is a shared object, not a chat transcript, so it lives in a real-time document model a group can edit simultaneously
• Build a defensible free wedge — a deterministic visa/passport-access engine that is genuinely useful standalone and feeds the paid planner

The build

• Nuxt 3 / Vue 3 SPA on the app tier, code-split per route, with a separate marketing/waitlist front end
• Firebase backbone — Auth for identity, Firestore for the live itinerary document model and real-time multi-user sync, Cloud Storage for media and OG assets
• LLM orchestration with structured-output prompting and schema validation, so every generated activity maps to a typed itinerary object the UI can render and edit field-by-field
• Retrieval and grounding pipeline — place resolution and enrichment against real-world POI data, then geo-clustering so each day is walkable rather than a scatter across the city
• Map-synced itinerary view — activities and map pins share one state tree; editing either side updates the other
• Visa and passport access engine — a rules dataset over passport × destination combinations returning visa type, cost, required documents and official advisories, plus destination ranking by accessibility and safety
• Import/parse path — paste text, a URL, or a PDF and the model normalizes it into discrete itinerary activities
• Public /explore surface — shareable itineraries as an SEO and virality channel

The results

• ⚡Itinerary in under 60 seconds — versus 8–20 hours of manual planning
• 🌍15,000+ trips planned across 120 countries (public figure, Nov 2025)
• 🚀Two Product Hunt launches — 103 upvotes and 247 comments on the planner launch, plus a second launch for the visa tool
• 💳Free-to-paid path established — the planner moved to a paid Unlimited tier without losing the top-of-funnel visa tool
• 📈Carried a full rebrand (Tern → Tilly) with no rebuild — brand, domains and OG surfaces swapped over the same app
• 🎙️Foundation extended to voice-first interaction and a taste graph, with a mobile app slated for early 2026

Tech stack

• Nuxt 3 / Vue 3
• Firebase Auth
• Cloud Firestore (real-time sync)
• Firebase Cloud Storage
• Structured outputs / JSON schema validation