How SayTrip Builds Trustworthy Travel AI on Sliplane

SayTrip turns conversations between travellers and travel advisors into structured, verifiable intelligence. We spoke with Co-Founder and CTO Michael Mayer about building trustworthy AI—and why a two-person company with deep Kubernetes experience chose not to run Kubernetes.

What is SayTrip, and what problem are you solving?

Michael: SayTrip is the intelligence layer for travel agencies and travel technology companies. It turns phone calls, emails, and chats between travellers and advisors into consistent, structured information.

It acts like an assistant that takes notes while the advisor focuses on the customer—capturing preferences, budgets, interests, and transport or accommodation requirements.

Unlike a generic summary, that information follows a defined travel-specific data model. Agencies can filter it, analyse it, and use it in their existing workflows.

“

The time Sliplane saves goes directly into improving SayTrip and working with customers.

How does it work during a consultation?

Michael: With the traveller’s consent, an advisor can record a consultation or upload an existing recording. SayTrip transcribes it, identifies the speakers, and extracts the relevant travel details using specialised models and prompts. Email is already being integrated, with more telephone and chat integrations to follow.

The system checks whether the traveller and trip already exist, then suggests creating or enriching their records. Every extracted detail remains connected to its original source.

If an email mentions a budget of €20,000 but a later conversation mentions €2,000, SayTrip flags the inconsistency instead of silently deciding which figure is correct. We also show whether the system is highly confident, confident, or uncertain about an extracted detail.

Advisors review the suggestions, especially uncertain ones. The result is a living traveller and trip profile that makes follow-ups and handovers much easier.

What is technically difficult about making that intelligence trustworthy?

Michael: A single general-purpose LLM may produce a plausible summary, but not necessarily trustworthy, reusable data. It does not inherently know which details matter to a travel business or how downstream systems need them structured.

We built that domain knowledge into a multi-stage pipeline. Even transcription is challenging: models must distinguish speakers, handle interruptions, and understand dialects such as Viennese.

We evaluate models from several providers. Some extraction tasks run in parallel for speed, while independent models perform consistency and logic checks.

If a conversation is about Sweden but an extraction suddenly introduces New York, the system can flag it. Every detail also links back to the relevant passage in the original conversation, so the advisor can inspect the evidence.

We maintain a benchmark suite with synthetic conversations and a human-reviewed reference set. Whenever a prompt, model, or pipeline step changes, we rerun it to catch regressions.

What have your first customers taught you?

Michael: Our first customers and design partners have been mid-sized travel agencies. Working closely with the people who use the product every day has helped us build the domain knowledge we lacked when we started.

The biggest lesson was that a trip is never planned in one conversation. Details evolve across calls, emails, and other messages, and remain useful even after the traveller returns. If one advisor is unavailable, a colleague should immediately understand the traveller and the state of the trip.

That shifted our focus from a single consultation to continuity across channels. Because we initially serve Germany, Austria, and Switzerland, data protection and GDPR are foundational design constraints.

What runs on Sliplane, and why did you choose it?

Michael: Almost our entire stack runs on Sliplane: the frontend, backend, databases, background jobs, internal tooling, and the infrastructure around our AI pipeline. The models run through specialised external inference providers, while recordings and files live in S3-compatible object storage provided through Sliplane.

My background is in DevOps and cloud-native infrastructure. I have managed Kubernetes clusters around the world and worked on migrations between AWS and Google Cloud, so I know both what those platforms make possible and how much operational overhead they create.

SayTrip is a two-person company, and I am responsible for the technical platform. Sliplane lets me deploy standard Docker containers and scale resources without taking on a platform team’s workload. Predictable costs and European infrastructure also fit how we are building the company.

What difference does it make day to day?

Michael: The clearest benefit is the feedback loop. With our automation on top of Sliplane, we can create a complete preview environment—including a seeded database—in around two to three minutes.

A staging or production release takes seven to nine minutes, including the build, tests, and deployment.

For a new service, I can connect its GitHub repository and let Sliplane build and deploy it without first creating a bespoke pipeline. Production still has proper CI and comprehensive tests, but experiments and internal tools do not need the same setup work.

With Sliplane, ongoing platform work is close to zero. I remain responsible for the application and its dependencies, but not the infrastructure underneath them. My experience would let me operate Kubernetes; it also helps me recognise that doing so would be the wrong use of our time.

Where can readers learn more about SayTrip?

Michael: Visit saytrip.ai or book a 20-minute walkthrough. We work directly with travel agencies, agency groups, and travel technology companies.