Why I started building Inora
Today I start on Inora’s retrieval system. At APG I assessed hundreds of survivor’s pension claims against policy and law. What that taught me about answers that must rest on a source, and why I am designing the retrieval from the ground up.
Written by
Co-founder, AI systems and product
Essay6 min read
I join Inora at a moment when a lot is already in place. On 2 September Inora went into production. On 4 September Vitala+ in Maastricht began working with it, the first care organisation to do so. On 15 September two regions followed: a network of care organisations in South Limburg, and in South-East Brabant the evening, night and weekend shifts in care for older people, when no physician is in the building.
My work will be the retrieval system: the part of Inora that, for each question, finds the passage the answer has to rest on. I am designing it from the ground up, directly on the language models, with no third-party layer in between. Below is why I want to, and why here.
What I did before
From October 2024 to March 2025 I assessed survivor’s pension claims at APG. Hundreds of them. Behind every claim was someone who had lost a partner or a parent. They were owed a decision that was right, and an explanation they could follow.
I tested every claim against policy and Dutch law. That sounds like administration. It is a discipline. A decision that does not point to a rule is an opinion, however good it sounds. The question was never whether an outcome seemed reasonable. The question was which rule it rested on, which version of that rule applied on the date in question, and whether a colleague with the same file and the same rule would reach the same outcome.
I also answered junior analysts’ questions there every day. The best answer I could give was rarely my own judgement. It was: look here, this rule, this paragraph. Then they could find it themselves next time, and check whether I was right. An answer that comes with where it was found makes the person asking stronger. An answer without it makes them dependent.
From those questions I wrote procedures, and the team adopted them. A procedure does not make a team cleverer. It makes everyone look in the same place. That turned out to be worth more.
A number never stands alone
Before that, in 2024, I built a genetic algorithm for SABIC’s container-ship routing. On sample data it found routes just over a fifth cheaper than conventional methods, and I presented the business case to SABIC’s management.
I give that number with its comparison and its data because without them it says nothing. Cheaper than what, and measured on what? Without that answer, “a fifth cheaper” is a loose figure. I will give every figure about Inora the same way: how much, against what, and over which period. A count tells you how much, not how good. If you want to know both, you have to measure both.
Why Inora
In care the same discipline is not a matter of tidiness. It is a matter of safety.
A nurse in doubt at night often has no one beside them to ask where it is written. The nurse decides. What they need is what I gave the junior analysts: their own organisation’s rule, the passage, and room to see whether it fits. Not someone else’s decision, and certainly not software’s. Inora does not diagnose, does not calculate a dose and does not decide how urgent something is. At APG I was the assessor. In care the care professional is the assessor, and Inora is where they find the rule.
Three things I found in Inora decided it for me.
First, the plan of 13 June. It put the foundation before any feature: Inora runs in an environment we control, shows the source with every answer and is tied to no single language model. A team that puts the source with every answer before there is a single user means it.
Then the architecture of 4 July. Each organisation gets a closed space of its own. Each document keeps its structure and its authority. Every question carries who is asking, and in which role. Those are exactly the things that decided an assessment: which document takes precedence, and whom a rule applies to.
And since 16 August Inora has recorded every intermediate step of every answer: which sources it used, and why. A wrong answer can be traced back to where it went wrong. Without that you cannot improve a retrieval system. You can only rewrite it and hope.
Why retrieval decides
A language model writes well. It writes an answer around whatever it is shown. Give it the right passage and the answer is as good as that passage. Give it the wrong one and it writes a fluent, confident, wrong answer, and no one can hear it in the tone. So the quality of an answer is decided before the model writes a word: in the retrieval.
In Inora that retrieval is no small task. In the sources Inora answers from, almost every question is half answered somewhere. The job is to find the passage that truly answers it, from the source that applies to this organisation, in today’s version. And to say there is nothing when there is nothing.
Too narrow is as wrong as too broad. On 3 September a new version made Inora too narrow, and within two and a half hours the previous one was back. The day before, the floor had asked for “client” instead of “patient”, the word long-term care uses, and twelve minutes later every answer used it. What Inora draws on, and in which language, decides what a care worker gets to read. That has to be correctable within hours, not weeks.
That is why I am building directly on the language models. A third-party layer between us and the model is a place where we cannot see why a passage was chosen, and where we cannot step in when it goes wrong. The June plan says Inora is tied to no single language model. For retrieval that means we decide how the search works, not the vendor of a layer in between.
How I will know it works
Not by how an answer sounds. By the passage under it: does it fit the question, is it the source that applies to this organisation, and would a colleague with the same protocol reach the same answer. And by what care workers themselves make of it, counted rather than estimated, by kind of question.
Inora is a care organisation’s knowledge system: what the organisation knows and has agreed, within reach of everyone who works there. Retrieval is the part of it I build.
At APG I learned that a decision is only as good as the rule it points to. In Inora an answer is only as good as the passage under it. Finding that passage, every time, is my work from today.