A booking platform rebuilt under deadline, for a team of two

The problem

View Retreats sold curated wellness and luxury stays, and for most of its first seven years it did so through a booking widget borrowed from someone else. In 2020 it signed a deal to source its inventory and bookings directly from a global travel partner, and set out to build that platform itself. A year later it was still a good way from launch.

The brief was unusually honest about why. “We totally underestimated the size and complexity of the API integration. This means that many things have not been planned or executed correctly.” The in-house team was two people: one with less than a year on the technology and, in their own words, thrown into the deep end, the other a week into the job. A third developer had just resigned. Pressure was mounting from inside the company and from the partner. And the catalogue had to grow from 650 properties in two countries to thousands across sixty.

They did not ask for another pair of hands. They asked for someone to get the shape right first, so they would not have to build it twice.

What we built

Bookings that tell the truth

The booking engine covers the whole journey: search, price, pay, confirm, amend and cancel. Because the travel partner is the authority on what a booking actually is, the site treats its own records as a copy and re-checks with the partner whenever anything changes, so a guest and the business never see different versions of the same reservation. When something goes wrong at the worst moment, a declined card or a room that sells out mid-checkout, the guest gets a plain explanation and a way back into the flow rather than an error code.

A site that stays usable while it works

The most urgent problem was not the booking engine at all. Importing a region of new properties locked the administration area, so nobody could do any other work while it ran. Those jobs moved into the background, where they can take as long as they need, with their own progress visible to the staff who care about them.

Images that look after themselves

Photographs are most of a retreat listing. They were moved onto fast worldwide delivery, resized on demand for each place they appear, and given an off switch the client could reach themselves if anything went wrong.

The hard part

The travel partner changed its image identifiers without changing the photographs, so the system reported thousands of new images that were not new. Someone had to open each pair and decide. In the client’s words, “in almost 80% of the cases, the image actually still exists, it just has a different image ID and thus, our system thinks it has changed.” They had been quoting the problem out to artificial-intelligence specialists.

It did not need a model. It needed the software to recognise a photograph it had already seen, which is a solved problem if you approach it the right way. That removed the entire class of false alerts and left the staff only the genuinely new pictures.

Making it run over the whole catalogue was the harder half. The first version took about twelve hours and then fell over, which is the worst of both outcomes. Seven weeks of work turned it into a job that steps over bad data instead of dying on it, works through the catalogue in parallel, and knows when it has finished. It now runs in about two hours and reaches the end.

What it changed

The platform launched on schedule in April 2021 with the booking and notification engine in scope, and the team stopped rebuilding work they had already done. The catalogue grew eightfold, from 650 properties in two countries to more than 5,300 across sixty, and stayed searchable. A daily manual chore that had been scoped as an artificial-intelligence project became something the system simply handles. The engagement outlasted the launch by two years and two major platform upgrades.

View Retreats closed in April 2025 after twelve years, citing the travel industry’s post-pandemic shift and the weight of much larger competitors. The platform ran the business until the end.