No longer dependent on your suppliers' AI choices.
Soev.ai is the AI platform for regulated markets. We bring AI into production at organisations with sensitive data or critical processes: tailored within weeks, and we measure whether it works.
Runs in your cloud or on-premiseISO 27001 certifiedDeployed at four ministries
Three layers stacked on top of each other. At the top the interface people work in: a chat with source citations and agents with a step for human approval. At the bottom the infrastructure it runs on: Dutch cloud, EU cloud, your own cloud or on-premise. Soev.ai is the layer in between, with four blocks: data and indexing, search, agents, and evaluation and optimisation. On the left the source systems that connect to it.
Runs where you wantinfrastructure
NL cloud
EU cloud
Your own cloud
On-premise
an isolated stack per customer · Kubernetes-native
control layer
Data & indexingparsing · OCR · chunking
Searchhybrid · reranking · permissions
Agentslog · human-in-the-loop
Evaluation & optimisationbenchmark · release gate
model gateway · guardrails · observability
ChatAgentsKnowledge bases
chat with your organisation
What help is there for an overburdened informal carer?
[1] Informal care policy rules.pdf[2] Social team meeting
Only sources you may access · logged
Ask a question about your documents…
agentsevery step in the log
New request for helpSearch the schemesProposalHuman approvalHandle the request
Deployed at more than 15 regulated organisations, including four ministries
Why we do this
By 2035, AI could support half of all work. We make sure organisations stay in control of that themselves: AI that demonstrably works, within the rules, in the place they choose themselves.
One platform. You choose how you use it.
The same blocks, stacked differently.
Two structures made of the same blocks. On the left a complete house for organisations: infrastructure as the foundation, on top of that connectors, models, the four core components and guardrails, with our workspace as the roof. On the right a pyramid for builders: infrastructure as the foundation, on top of that the four core components, above them models and guardrails, then one API, and your own software at the top.
for organisations
You want to use sovereign AI across your whole organisation.
A complete house
All the blocks, from foundation to roof, connected to your own systems. We build it and run it, in the cloud you trust.
For municipalities, executive agencies and knowledge organisations.
Four things we take care of in every implementation.
[01]
Always in control
Your cloud or on-premise
Open source
No lock-in
Four places where soev.ai can run: a Dutch cloud at one of our cloud partners, a European cloud, your own cloud or on-premise on your own servers.
open-source frameworkan isolated stack for every customer
Dutch cloudAt a Dutch cloud partner.
EU cloudIn a European cloud.
Your own cloudIn your own cloud environment.
On-premiseOn your own servers.
Runs where you want.
The platform is complete and modular: interface, gateway and models can be swapped for your own. The framework is open source, and your data and whatever is set up specifically for you remain yours.
Kubernetes-native, Haven+ compliant. An isolated stack per customer.
[02]
Compliant
Your data stays yours
Guardrails
Auditable
A question containing a name and a citizen service number. PII masking is switched on here: the personal data is masked before the question reaches the model. Below it three checks: no prompt injection found, two items of personal data masked, and the answer is backed by sources. The step is recorded in the log.
Guardrails on every call
question
My client name, BSN BSN, is asking for respite care. What is he entitled to?
Prompt injectionnothing found
PII maskingoptional · 2 masked
Grounding checkbacked by 2 sources
recorded in the log
The log of one agent run as a timeline: question received, sources searched, proposal drafted, paused for human approval, approved, carried out and recorded.
Log of one runimmutable
[01]
Question received
What the agent was given.
[02]
Sources searched
What the agent saw.
[03]
Proposal drafted
What the agent decided.
[04]
Waiting for approvalPaused
A member of staff checks the proposal.
[05]
Approved and resumed
Who approved it is recorded.
[06]
Carried out
What the agent did.
Your data stays yours, and every step can be checked.
Your data stays yours. Never trained on your data, in the cloud you choose.
Permissions follow the source. Enforced on every search.
Guardrails. Prompt injection detection and a grounding check. PII masking if you want it.
Schematic chart of quality and consumption per task. A frontier model sits at the top right: high quality, high consumption. A small open model starts lower and reaches the same quality in four steps, with much lower consumption.
1Small open model, not yet tuned.
2Search and context tuned to your data.
3Prompts optimised automatically with DSPy and GEPA.
4Tested on your benchmark, and only then live.
Schematic. Every version is measured before it goes live.
We tune the harness to your data and your task: quality up, consumption down.
We optimise automatically, and a new version only goes live if it scores at least as well.
Timeline in five steps: introduction, agreeing the benchmark, building phase 1, live and measuring, and then managed service. In production within weeks.
[01]IntroductionWhat do you really want to solve?
[02]Agree the benchmarkThe yardstick, up front.
[03]Build phase 1A production-ready first version.
[04]Live and measuringDoes it score on the benchmark?
[05]Managed serviceWe run it, you take over step by step.
