60% lower infrastructure spend through smarter supplier routing.
A travel aggregator's infrastructure bill grew in lockstep with its traffic, because every customer search fanned out to every supplier API, whether or not that supplier was likely to return a bookable result. Plexe built a real-time routing model that decides which suppliers are worth calling for each search, cutting infrastructure cost by 60% and making responses faster.
The challenge
- Every search hit every supplier API
- Long-tail suppliers consumed budget but rarely converted
- Infrastructure costs scaled linearly with traffic
- Slow response times hurt the experience
What we built
- ML models that score and rank suppliers in real time
- Queries routed to high-performing suppliers only
- Infrastructure spend decoupled from supplier count
- Faster responses from fewer, better API calls
Results
- -60% infrastructure cost
Why they came to us
An aggregator's value is breadth: the more suppliers it can search, the better the prices it can show. But breadth has a cost. Each search triggered a call to every connected supplier, so infrastructure spend scaled directly with traffic, and every new supplier made each search more expensive.
Most of that spend went to the long tail. A small set of suppliers produced the bulk of bookings, while many others were queried on every search and rarely converted. Waiting on all of them also slowed results down, which hurt the customer experience the breadth was meant to improve.
Goals and success criteria
The brief was to cut infrastructure cost substantially without hurting what customers see. That meant routing decisions had to happen in real time, inside the search request, and the system had to keep calling the suppliers most likely to return the best results. Faster response times were a requirement, not a nice-to-have.
How it works
Plexe trained ML models that score and rank every supplier for each incoming search, using signals like route, dates and each supplier's history of returning results that convert. Searches are then routed only to the suppliers worth calling, and the long tail is queried only when it is likely to add something.
The models sit directly in the search path and return a decision in real time. Fewer, better API calls mean lower spend and quicker responses, and adding a new supplier no longer raises the cost of every search.
Results and what they own now
Infrastructure cost fell by 60%, and spend is no longer tied to the number of connected suppliers. The routing models run in the aggregator's own stack and are maintained by its own team, so they can retrain as suppliers and traffic patterns change.