AI product engineering · WanderBu
Designing WanderBu for high-confidence AI travel itineraries
A useful itinerary needs more than fluent recommendations. WanderBu separates understanding, research, discovery, evaluation, and repair so every journey can be checked before it reaches a traveller.
- Written by
- Imajin Editorial Team
- Published
- Reading time
- 9 min read
In brief
What this article covers
- Gemini provides reasoning and synthesis across a structured generation workflow.
- Google Places API grounds candidate places and location context in real-world data.
- Retrieval-augmented generation adds curated travel knowledge beyond general model context.
- Nine explicit stages create opportunities to score, check, and repair a journey before presentation.
Product premise
Travel planning has an information problem—and a confidence problem.
Travellers can already find an overwhelming number of destinations, reviews, maps, videos, and generated lists. The harder task is deciding whether those fragments form a journey that can actually be taken by a specific person or group.
WanderBu is designed around confidence rather than recommendation volume. It reasons about intent, constraints, booked commitments, transfers, pacing, luggage, mobility needs, route risks, and alternatives. The aim is not to make the longest itinerary. It is to make a useful one whose details support real decisions.
Intelligence stack
Combining model reasoning with places and retrieved knowledge.
Gemini is the reasoning and synthesis layer within WanderBu. It helps interpret natural-language travel intent, connect constraints, compare possibilities, and turn structured findings into a coherent plan. We do not treat the model as the sole source of travel truth.
Google Places API supplies place and location context for discovery and grounding. It helps the system work with real destinations and points of interest rather than relying only on names produced during generation.
We also retrieve additional curated knowledge through retrieval-augmented generation, or RAG. This gives the generation process relevant domain material at the point it is needed and makes specialised guidance available beyond the model’s general context. Each source plays a distinct role: reasoning, real-world place context, and additional travel knowledge.
Nine-stage process
A staged pipeline creates room to verify and improve.
A single prompt-to-itinerary response compresses too many decisions into one opaque step. WanderBu instead moves through nine stages. The stages are designed as a quality system: later steps can inspect, challenge, and repair the work produced earlier.
- 01
Understand
Interpret the travellers, purpose, preferences, constraints, dates, group needs, and existing commitments.
- 02
Research
Identify the questions that must be answered for this destination, route, season, and traveller profile.
- 03
Acquire knowledge
Retrieve relevant curated guidance and supporting context through RAG.
- 04
Discover
Find suitable places and experiences, using Google Places API where location grounding is required.
- 05
Score
Evaluate candidates against relevance, confidence, feasibility, traveller fit, and journey quality.
- 06
Organise
Sequence selected elements into realistic days, routes, transitions, and decision points.
- 07
Check
Inspect the assembled itinerary for gaps, contradictions, impractical pacing, and missing essentials.
- 08
Repair
Correct weak or conflicting details and introduce safer alternatives or backup plans.
- 09
Present
Translate the verified journey into clear, useful guidance a traveller can act on.
Travel context
The itinerary is only one part of being ready.
A route can be geographically correct and still leave a traveller unprepared. WanderBu accompanies the plan with practical destination context, including visa information, the local vibe, emergency details, common travel phrases, connectivity guidance, booking priorities, comfort considerations, route risks, and backup options.
These details are organised around the journey rather than presented as an unrelated destination guide. The intention is to help travellers understand not only where to go, but what to prepare, what may feel unfamiliar, and what to do if conditions change.
What comes next
Planning should remain useful when the journey begins.
Travel does not stop being complex after an itinerary is generated. Bookings change, groups coordinate, conditions shift, and useful discoveries happen along the route. WanderBu is being designed as a continuous travel-confidence layer rather than a one-time document.
An accompanying WanderBu mobile app is coming soon. It is intended to carry the shared itinerary, practical guidance, booked commitments, emergency context, and relevant journey information into the travel days themselves. Availability and release details will be announced when the app is ready.
Quality principle
Confidence must be earned, not implied.
Travel information can change, and model output can be incomplete. WanderBu’s architecture therefore treats checking and repair as first-class product stages. It is designed to expose useful cautions and alternatives rather than hide uncertainty behind polished prose.
The product will continue to evolve through real usage, source evaluation, internal quality scoring, and feedback from travellers and travel operators. The durable advantage is not one model response; it is the system around that response.