Skip to content
FLX
DE
Menu

Build

AI integration

Putting language models into existing processes: a fixed output format, visible cost, a swappable provider.

What you get

  • Versioned prompts with a fixed JSON output format
  • The cost per operation, known before the feature goes live
  • An evaluation showing the cases where the model is wrong
  • A swappable provider, and a way back to the previous process

Approach

How we go about it

We build LLM features with versioned prompts and a fixed JSON output format, so everything downstream can rely on it. An evaluation shows the cases where the model is wrong, and the cost per operation is known before the feature goes into production. The provider stays swappable, which makes changing one a configuration question. So far we have built this in two forms: systems in which several agents work through a process on their own, and a language model of our own that answers questions about one particular product.

Build in the log

Entries of this kind from every system we have worked on.

  1. Build

    Built the receipt pipeline, implementing RFC 6376 DKIM verification ourselves rather than trusting a library to get exactly that right.

  2. Build

    Moved every tenant onto one Kustomize base with per-tenant overlays, so a release ran the same way everywhere.

  3. Build

    Started the multi-tenant platform: API gateway and the first backend services.

To the log

More services

The other services

  • Build

    Software & platforms

    Applications and multi-tenant platforms for companies whose processes no longer fit a standard product.

  • Data

    Data & analytics

    Bringing figures from separate systems together, reconciling them and making them fit to report on — including the places where two systems give different numbers.

  • Operations

    Operations & infrastructure

    Your system runs and somebody is responsible for it: deployment, databases, monitoring, provider migrations.

  • Build

    Automation

    Letting software do the recurring manual work — with a dry run, an approval and a log.

  • Advisory

    Advisory & architecture review

    A second pair of eyes on architecture, security, and which part of the plan can be dropped.

  • Training

    Training & enablement

    Your team takes the system over and carries on developing it.

We will tell you whether this is the right service.

Describe in a few sentences which system this is about and what it has to do. We will tell you which service fits and what a first step looks like.