ZeniTech · Case studies · Travel · Français

Zeniva Travel: an online travel agency, and an assistant that depends on no single AI provider.

Zeniva Travel sells high-end trips to American travelers, through independent travel advisors. We built the website, the advisor workspace and Lina, the conversational assistant that answers clients at any hour.

Next.js 16React 19TypeScriptSupabase / PostgresVercelAmadeusDuffel

Disclosure: Zeniva Travel, ZeniPay and ZeniCorp are companies in the same group as ZeniTech, founded by the same person, Alexandre Blais. They are not outside clients: we built these platforms for ourselves, and we run them.

Where the project stands

The site is live. Search, the advisor workspace and the assistant work. Influencer commissions are calculated on the server; advisor commissions are still calculated in the interface and need to move to a server-side ledger. We claim no booking volume here.

What we built

A complete travel website and a workspace for advisors: 163 pages and 173 API routes, in Next.js 16, React 19 and TypeScript, on a Postgres database hosted by Supabase, with row-level security policies and role-based access control.

Flight and accommodation search runs through the Amadeus and Duffel APIs. The site also works as an installable app, with notifications, and each partner agency can have its own configuration of the assistant.

The technical decision that matters: no single provider

Lina does not talk to a single language model. Every question goes through a cascade: if the first provider does not answer, the request moves to the next one, and so on across five options, including Groq, Anthropic, Google and OpenAI. When a provider retires a model, the assistant falls back on its own to a known model.

This is an architecture choice, not a detail. An assistant wired to a single provider goes down with it, and changes price when that provider changes its prices. The cascade costs a little more code up front and removes that risk for good.

Every exchange is stored with its human rating and, when needed, the corrected answer. That is what makes it possible to measure the assistant’s quality instead of assuming it.

What it teaches you if you are having software built

If your project uses a language model, ask what happens the day that provider is down or doubles its prices. If the answer is “we wait,” that is a risk you are buying along with the software.

Your project

We build yours with the same rules.

Code repository in your name from the first commit, standard stack, fixed price per phase. Free 30-minute discovery call, in English or French.

Other case studies