AI engineering solutions we build.

• Knowledge assistant on your own documents

  Answers with sources, taken from your own documents.

  Retrieval returns the passages together with their sources, and the access rules of the source system apply to every answer. We build an evaluation set of real questions first and rerun it after every change.

• AI workers for back-office roles

  Software that takes over a repeatable task, with a person checking.

  Each worker has one defined role, scoped tools and an approval step where it matters. Every action is logged, so you can review what it did and why.

• Voice agents

  Phone and web agents that hold a conversation and hand over to a person.

  We design the call flow, the hand-over rules and the fallbacks, and test with recorded calls before go-live. Transcripts and call outcomes are logged for review.

• MCP servers for internal systems

  Read-first tools that let agents use your ERP, bookkeeping, SharePoint or TMS.

  Each server starts with a small set of read tools. Write actions follow only after review. Scoped service accounts and an audit log limit and record what an agent can touch.

• Governed coding agents

  Coding agents for your developers, with rules, review gates and budgets.

  Requests are routed through a gateway, so spend and model choice are controlled centrally. Repository rules and mandatory review gates keep every agent change inside the normal pull request flow. Omnigent from Databricks is an option for policies above the agents (open source in alpha, managed version in Beta).

• Talk-to-your-data on Databricks

  Business users ask questions in plain language, on curated data.

  Genie works on curated Unity Catalog semantics: agreed definitions, certified tables and access per group. We test it against the questions your teams actually ask.

• MLOps platform

  Models move from notebook to production through a governed path.

  A registry in Unity Catalog, training and deployment pipelines as code, serving and drift monitoring. An evaluation gate stops a release when quality drops, and your team operates the result.

• Self-hosted LLM

  An open-weight model in your own environment.

  We size the hardware, serve the model behind an API and add monitoring and access control. Your data stays in your environment.

• Fine-tuning on your data

  Adapt an open-weight model to your tasks and vocabulary.

  The base model is open-weight under a permissive licence, checked per model. We measure before and after on your tasks, and the weights stay in your environment.

• Model choice for Dutch

  Find out which model works best on your Dutch tasks.

  We benchmark candidate models on your own Dutch tasks and compare quality, cost and hosting options. You receive a decision table you can defend.

• Small models at the edge

  Quantised models on factory or vehicle hardware.

  We quantise a model to fit the device, then measure accuracy and latency on the real hardware. Updates roll out through a controlled pipeline.

• Private model hub

  One governed registry for models and datasets.

  EU storage, single sign-on and audit logs. Teams find approved models in one place, with the licence and the evaluation results next to each.

Bring your own model: switch models without rebuilding the harness. How swapping a model works

• AI gateway and cost control

  One control point for access, budgets and logs.

  Built on Unity AI Gateway: one service per use case with its own budget and owner, and usage dashboards per team.

• Agent evaluation and guardrails

  Test sets and guardrails before an agent reaches users.

  We build a test set from your own cases. Judge models are checked against human ratings, a regression gate runs in the pipeline and tracing shows every step an agent took.

• LLMOps and evaluation

  Prompts, models and answer quality under version control.

  We build evaluation sets from your own cases, check judge models against human ratings and run a regression gate in the pipeline. Tracing, inference tables and budgets per use case complete the audit trail.

• AI Act compliance kit

  Register, classify and document each AI system.

  A model register, the risk class, technical documentation, logging and human oversight per system. We start from the public AI Act checker and work through your systems one by one.

• Anonymisation and synthetic data

  Safe test and training sets from sensitive records.

  We remove or replace identifiers and check what can still be re-identified. Synthetic data is validated against the real distribution before anyone uses it.

• Document AI

  Typed-over documents become structured data.

  Uncertain fields go to a person, with an audit trail. We measure accuracy on your own document types before production.

• Transcription and summaries

  Meetings and calls become searchable text, on your own infrastructure.

  Speaker labels, and summaries that link back to timestamps in the recording. Everything runs on infrastructure you control.

• Translation pipelines

  Dutch and English with your own terminology.

  A glossary enforces your terms and a review step catches drift. We measure quality on your own texts before roll-out.

• Forecasting with time-series models

  Demand, energy or capacity forecasts, back-tested against today's method.

  We back-test on your history against your current method, so you see whether a foundation model adds anything. Forecasts are served with their uncertainty ranges.

• Vision inspection

  Detection and segmentation on a production line or camera.

  We label a sample set, train and evaluate on your images, and test on the real camera set-up. The model licence is checked before delivery.

• Predictive maintenance

  Early warnings from sensor and machine data.

  We build the telemetry pipeline, train on your failure history and alert maintenance with the reasons. Models are monitored for drift after go-live.

1. 01

   Scope one process

   One process, one reader or one asset class, with a measure you can compare with today.

2. 02

   Build with evaluation

   An evaluation set from your own data exists before the first release, and every change reruns it.

3. 03

   Run beside the team

   The system runs next to the current way of working, so people can compare and correct it.

4. 04

   Hand over

   Code in your repositories, a runbook and a handover session.