About

We got tired of guessing.

SebyAI was built by a team that shipped LLM features before the tooling existed, and felt every gap.

Our story

It started with a bad week.

In early 2023 our founding team was running a support assistant for a mid sized logistics company. It worked well for four months. Then, over about nine days, answer quality fell off a cliff, and nobody could say why. The model had not changed. The prompt had not changed. Something upstream had.

It took eleven days and a great deal of manual log reading to find it: a retrieval index had silently stopped refreshing, so the assistant was answering current questions with three month old policy documents. Every dashboard was green the whole time. Latency was fine. Error rates were fine. The system was confidently wrong and perfectly healthy.

That is the gap we build for. Traditional monitoring tells you whether a service responded. It has nothing to say about whether the answer was any good. SebyAI exists to close that distance, and to make the closing of it boring and routine rather than an eleven day archaeology project.

How we work

Read only by default

The collector observes. It does not sit in your request path and it cannot take your product down. If SebyAI fails, your application does not notice.

Your data stays yours

Self host the collector and nothing leaves your network. On our cloud, prompt and completion bodies are redacted at the edge before they reach storage.

Boring on purpose

We ship small changes often and we publish our incidents. Observability tooling that surprises you is not doing its job.

The team

Nineteen people, six timezones.

We are deliberately small. Most of the team has carried a pager for a system that made decisions on someone else behalf, which shapes what we choose to build and what we refuse to.

Engineering is fully remote. Support is covered by engineers on rotation, not a separate tier, so the person answering you can usually read the code.

Small team reviewing work together in a bright meeting room

Come see whether we are right.

We would rather show you the product against your traffic than talk about it.