AI engineering: document automation, LLMs and agents

01 / Problems

Problems we solve

Typical starting points for an AI engagement.

Paper and PDFs run the back officeBack-office processes are document-heavy and manual, and the volume does not fall.
LLMs under strict rulesYou want to use LLMs, but data classification, sovereignty and audit expectations are strict.
The data is in separate systemsThe data an AI system needs sits in different places, so a model has nothing reliable to work from.

02 / Approach

How we deliver

How an AI system gets to production with us.

One process first
We pick one document type or one question, so the result can be measured against what happens today.
Your tenant, your data
The pipeline runs in your own Azure tenant. We do not send your data to external AI services without your written agreement.
Human check and evidence
A review step and an audit trail, and accuracy measured on your own documents before anything goes live.
Same pipelines as the platform
Reviewed pull requests, dev to production deployment and a run book, so your team can operate it.

AI harnesses and frameworks

One of our main expertises is the engineering frame around models and agents, so they run reliably in production and your teams ship new AI use cases faster.

One processMeasured against today
OrchestrationTool access via MCP
ModelIn your tenant
GuardrailsInput and output checks
Tracing and evaluationRegression gate

Agent harnesses and orchestrationTool-use loops and multi-agent graphs, for example with LangGraph, and tool access through MCP.
Evaluation frameworksTest sets built from your real tasks, regression gates on every change, and LLM-as-judge with human review.
Guardrails and policy checksInput and output checks and access rules that run on every call, not only in review.
Observability and tracingTraces of every step with tools such as MLflow and OpenTelemetry, so failures can be found and explained.
Cost and latency controlLimits, caching and model routing, so cost and response time stay within what the use case needs.
Reusable internal frameworksA shared frame and templates, so a new AI use case starts from working parts instead of from zero.

MLflow
OpenTelemetry
Azure

Where AI projects stall

Prototypes that do not survive production. A demo works on the happy path. Without control over tool use, retries, limits and monitoring, it fails on real traffic.
No way to measure quality. Nobody can say whether a change made the system better or worse, so releases rest on opinion.
Every team reinvents the plumbing. Each team builds its own orchestration, logging and checks, and none of it is reused.

What you get

A harness in your repositories, with orchestration, tool access and guardrails
An evaluation suite with a regression gate in your delivery pipeline
Tracing and cost views connected to your observability stack
A reusable template and a short guide, so your teams add use cases on the same frame
A run book and a handover to your team

Your model, swappable

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

Cost and token optimisation

The harness is where cost per task is measured and reduced, because every change runs against the evaluation set before it ships. Cost and token optimisation in depth

All AI engineering solutions

Governance for AI

One gateway for every model, agent and tool

Teams, agents, coding toolsWhoever calls a model
Unity AI GatewayPermissions, budgets, limits, policies, logging
Models, MCP servers, toolsDatabricks-hosted and external

AccessUnity Catalog privileges decide who may use what.
Budgets and limitsSpend and request rates have a cap per use case.
PoliciesRules apply to requests and responses.
LoggingEvery call lands in Delta tables for audit.
RoutingTraffic splits across models, with failover.

How we set it up

The gateway and its policies are defined as code and deployed by pipeline, never clicked together.
One service per use case, each with its own budget and an owner who answers for it.
Inference tables with a retention period and access groups, so logs are useful and not a liability.
System-table dashboards for tokens, cost and latency per team.
Coding agents such as Claude Code are routed through the gateway, with repository rules and review gates on top.
Fallbacks are tested before go-live, by switching a destination off on purpose.
Smart Routing stays out of production while it is in Beta.

Generally available. Smart Routing is in Beta. Last verified: 10 November 2026. Unity AI Gateway documentation

Databricks
Unity Catalog
Unity AI Gateway
Delta Lake

ML and LLM operations

Models and LLM apps that stay in production

We build the path to production on Databricks: registry, pipelines, an evaluation gate, serving and monitoring, for classical models and LLM applications alike.

ML and LLM operations in depth

Voice agents

Voice agents that answer the phone

Built with LiveKit for real-time audio and ElevenLabs for speech. We have built many of them.

Answer and route calls to the right team or person
Intake and booking, with the details written to your systems
Multilingual, including Dutch, with a handover to a person

Problem, approach and expected outcome

VA

Voice agents for inbound calls

The problem: Phones ring out of hours or queue at peak, and staff spend time on routine questions and on passing callers to the right person.
What a voice agent does: An agent answers the call, understands the request, routes it to the right team, takes intake details and books appointments. It speaks several languages, including Dutch, and hands over to a person when it is unsure.
Expected outcome: Aims to shorten waiting time and take routine calls off your team. These are expected outcomes, not results: call volume, languages and integrations decide what is realistic, which is why we pilot on one real call flow.
How we deliver: We have built many voice agents with LiveKit and ElevenLabs. We measure how well it works on your own call flow.

Where call content is sensitive, we can run the language model on infrastructure you control. ElevenLabs is a hosted service, so a fully self-hosted voice agent swaps in open speech models, after we have evaluated whether they are good enough for your languages.

Sovereign and self-hosted AI Talk to the demo voice agent

Inbound callIllustration

Live

Hello, I would like to book an appointment.
Of course. Which day suits you?

Routed toBooking

Deliverables

What you get and how to start

Each starts with a conversation about your scope.

Typical duration: 2 to 3 days

Data & AI quick scan

In 2 to 3 days we map how data and documents flow today and rank the three AI or data use cases with the best value for effort.

What you get

Current state of systems, data and manual work
Three use cases ranked by value, effort and risk
Roadmap with a pilot proposal
Session with management to discuss the result

Ask for a quote

Typical duration: 3 to 4 weeks

AI document automation pilot

We turn one document stream into structured data with a human check, running in your own Azure tenant.

What you get

Working pipeline for one document type
Human review step and audit trail
Accuracy and time-saved report
Handover and run book

Ask for a quote

All engagement models

Proof

Related case studies

Outcomes are quoted from our case studies. Your result depends on your data, your teams and your starting point.

All projects

Knowledge

Take the check

A score from 0 to 100 and a next step, in about three minutes.

All six checks

FAQ

Questions about this service

What is an AI harness?

The engineering frame around a model or agent: orchestration and tool access, evaluation, guardrails, tracing and cost control. It lets an AI system run reliably in production and lets your teams add new use cases on the same frame.

Where do the models run?

In your own Azure tenant for the document automation pilot. If an external AI service is the right tool, we only use it with your written agreement.

Do we need a data platform first?

Not for a first pilot on one document type. For use cases that combine several systems, a governed platform is the foundation, and the quick scan tells you which case you are in.

How long does it take?

The quick scan takes 2 to 3 days. A document automation pilot takes 3 to 4 weeks.

What if the use case does not pay off?

The quick scan ranks use cases by value, effort and risk, so you can decide not to proceed. The pilot ends with an accuracy and time-saved report, so you see evidence before investing in production.

AI engineering that starts from one real process.

Problems we solve

How we deliver

One process first

Your tenant, your data

Human check and evidence

Same pipelines as the platform