Skip to content

We show nothing you cannot check yourself.

Systems grounded in your sources, wired into your tools, that stop at a human when the decision calls for one.

What we build

Five building blocks. The process decides which ones, never the other way round.

  1. Knowledge and document retrieval

    Answer from your documents and internal systems, cite the source, and stay quiet when it is not enough.

    Chunking, embeddings, reranking, citations, access permissions, evaluation

  2. Multi-agent workflows

    Split a long job across specialised components that research, decide, act and check the result.

    Orchestration, queues, state, human approval, retries

  3. Multimodal systems

    For what text alone cannot handle: reading image-based documents, searching a visual catalogue in plain language.

    Automatic image captioning, text-to-image retrieval, image document parsing

  4. Deployment and infrastructure

    Hosting is chosen for the sensitivity of the data and what your team can run, never for a preferred vendor.

    Private instances, European hosting, Docker, PostgreSQL, self-hosting

  5. Governance and reliability

    The gap between a demo and a system you can rely on is decided almost entirely here.

    Grounding, evaluation, logging, confidence thresholds, exceptions, GDPR

Three systems that ran in production

Not slides. Real data, real users, real failure cases.

Answering from knowledge the company already holds

A system that ingests documents, SharePoint content and web pages, then retrieves the relevant passage along with the one before it and the one after. Deployed on e-commerce product questions and first-level legal research.

  • Reranking
  • Private instances in Europe
  • Anti-hallucination guardrails

Letting the customer describe or show what they are after

A chat adviser that understands a sentence or a photo, and finds the item in a catalogue with no text in it. The images are captioned automatically, and the search runs against those captions.

  • Automatic image captioning
  • Text-to-image retrieval

Nothing goes out before a human has seen it

Four specialised agents pass the work between them, and the gap between the draft and what was actually sent is what the system learns from. The setting was outbound sales; the architecture fits anywhere an action needs approval before it goes out.

  • Self-hostable
  • Human approval queue
  • Learning loop
  • Open source

This work predates the company: a previous role, and products we built ourselves. It is not client work delivered here.

Who you actually get

One of us brings five years of software development, two of them full time on document retrieval systems in production. He has built two AI products end to end, one of them public and open source.

The other brings fifteen years of selling and deploying enterprise software across Europe, Asia and the United States, now applied to AI workflows.

No juniors, no subcontractors, no handover to an account manager in month three. You get both founders, from the diagnostic through to delivery.

Ask us to walk through one.

One of the three is public and open to inspection. Check it rather than take our word for it.

Book the diagnostic