(03) — Case Study · Youmi (co-founded)
Youmi Travel
An AI travel-planning product built from scratch with a planning-first agentic pipeline.
Role
Co-Founder & Founding Engineer
Company
Youmi (co-founded)
Timeline
2024 – 2026
Focus
AI Agents · Flutter · NestJS · SSE
0 → 1
full product, built from scratch
7
pipeline stages
2
platforms: iOS & Android
Overview
Youmi is an AI-powered travel-planning product I co-founded and engineered end to end: a Flutter mobile app, a NestJS backend, and a planning-first AI pipeline that produces executable, ready-to-use itineraries instead of raw attraction lists. Built independently, outside my role at Visa.
Gallery
Architecture
- Intent extraction
- Place search
- Event enrichment
- Route optimization
- Itinerary generation
- SSE streaming
- User revision loop
Challenge
Most AI travel tools dump a list of attractions and call it a plan. A real itinerary has to respect opening hours, travel time between places, meal timing, and the user's actual intent — and it has to arrive fast enough to feel interactive, not like a batch job.
Approach
I architected a planning-first pipeline that separates reasoning from rendering: extract intent, retrieve and enrich places via Google Places and routing APIs, optimize routes, generate the itinerary, then stream results progressively. Keeping planning upstream of generation cut LLM token cost and made outputs verifiable at each stage.
Delivery runs on SSE with Redis and BullMQ for progressive streaming — the itinerary builds live on the user's screen as stages complete. A multi-agent orchestration layer classifies intent and routes requests dynamically, and a revision loop lets users adjust any part of the plan without regenerating the whole trip.
The client is a cross-platform Flutter app with real-time streaming and map-based interactions.
Outcome
A working 0→1 AI product: executable itineraries with live streaming UX, resilient API integrations with caching, validation, and fallbacks. The architecture patterns — planning-first pipelines, SSE streaming, agent orchestration — carried directly into my current work building autonomous travel operations systems.