SproutRoute — family trip planner
SproutRoute is a family trip planner that brings day-by-day itineraries, route maps, packing, and safety guidance into one plan.
1
Proof
21–30 s
First day of the plan on screen
Nine production runs of three trips (Orlando + Miami, Lisbon, Rome), Nov 8, 2026. Model generation is most of it.
69 ms
Family-fit judgment with a decision model
Jev, median of 303 labeled places; Claude Haiku took 1,140 ms at the same accuracy. Nov 8, 2026


2
The iOS app
Describe the trip, get the plan
A parent types the trip the way they would text it. The app returns a summary, the forecast and a day-by-day itinerary. Screens from the November 2026 TestFlight build.

The request, in plain words. 
Trip summary and forecast. 
Day 1, with a nap break built in. Change it, pack for it, check the rules
Each stop can be moved, swapped or rejected. The packing list and the car seat and pet cards come from the same trip details. The car seat card shows the app's own summary and points to the official source.

Editing a stop. 
Packing checklist. 
Car seat and pet cards, as the app shows them.
3
The work
6 chapters
3.1
The problem
A list of attractions isn't a trip plan. Families also have to decide which cities to visit, how to move between each day's stops, what to pack, and what to check before leaving.
3.2
Why it mattered
A plan families can rely on needs clear loading states, recoverable failures, and an honest line between AI suggestions and sourced information.
3.3
What I did
AI for suggestions, rules and APIs for facts. AI proposes the itinerary; rules produce packing lists; dedicated services supply weather, place details, and map directions.
Plans come from a curated database: 547 destinations and about 10,400 family attractions, 7,500 of them matched to Wikidata for coordinates that are free to show on any map.
Moved packing from AI to rules after the AI output proved inconsistent.
Put safety highlights on the main plan instead of a separate tab. The full guidance stays one click away.
Plan cities in parallel. Two city plans run at once, and packing and safety start without waiting for the itinerary.
Moved the family-fit judgment for each attraction to a decision model (TypeSafe's Jev). It answers in about 70 ms, and Claude Haiku takes the places Jev is unsure about.
I took SproutRoute from the first idea through design, implementation, and launch, and I operate it.
3.4
What went wrong
A Japan demo exposed missing attractions, slow generation, and maps that showed a city instead of the day's route. I traced the browser requests and production timings, repaired the failure paths, and rechecked the full trip on desktop and a narrow screen.
Evals found what the demo did not. The attractions database had been ignored since July: the query that loaded candidates returned no rows for any city, so the model planned from memory. That query is fixed. A live plan also put the Museum of Flight fifteen minutes before closing, which led to a program check on opening hours.
Venue and safety claims are still not checked by a person. Scoring ten full plans by hand is the next step.
3.5
Stack
React and Express on Railway, Supabase, Google Maps and Places, Wikidata, and weather APIs. Gemini 3.8 Flash writes itineraries, with GPT-6 Luna and Claude as fallbacks; TypeSafe's Jev judges family fit.
3.6
A route for each day
Named attractions stay on the map even when coordinates are missing, and Google Maps connects the day's stops in itinerary order. The map answers "how do we get between these?" instead of showing a city pin.
Build notes: model choice, reliability, architecture
Model choice against the real workload
I compared three models on the production prompt: Tokyo trips of three and six days for adults, and a three-day Kyoto trip with children. Two passes gave six samples per model. GPT-5.4 nano averaged 28.1 s and produced an invalid activity reference and a repeated activity. Gemini 3.8 Flash averaged 13.8 s and met the requested activity coverage and reference checks, so it runs itineraries. Gemini 3.5 Flash Lite averaged 10.7 s but returned fewer than four activities on some days.
The benchmark checks output structure and coverage. It doesn't establish venue accuracy, opening hours, or safety quality.
Reliability fixes after the Japan demo
API-origin fixes and explicit retry and edit actions replaced blocked requests and silent regeneration after a stream error.
Named attractions stay on the map without coordinates; the route follows itinerary order.
I restored the protected metrics dashboard's sign-in and metric mappings, then used stage and model timings to find the slow calls.
Architecture
Parse and review: the parser extracts travel constraints, and travelers confirm or reorder a multi-city route.
Plan and stream: the backend resolves destinations, retrieves weather and cached attraction candidates, and streams itinerary results.
Prepare: rules produce packing items; separate requests supply safety guidance and place details.
Measure and recover: protected metrics record model latency, trip stages, and errors. Partial results stay visible when a later stage fails.
Checks
533 unit and integration tests and 73 browser tests passed for the September 9 release. The 25.0 s figure is one observed run, not a response-time guarantee.
4
Contact
Hiring for a product area?
I am looking for Principal / Staff IC ownership of a product area.