Michiel Dondorp

Predict-IT

Additional Speakers

Mark Rienks

From Feedback Backlog to Better Requirements: TenneT’s Case Toward Governable AI

Bio.
Michiel Dondorp werkt sinds 2016 aan de combinatie van Design, AI en Systems Engineering in en rond infrastructurele projecten. Hij heeft een basis als werktuigbouwkundig ingenieur en een technische achtergrond in computational design, softwareontwikkeling en AI. Als Product Lead bij predictIT richt hij zich op AI-ondersteunde Systems Engineering  binnen energie, water en infrastructuur.

Abstract.

Organisations that write and manage requirements at scale run into this bottleneck: processing large amounts of valuable feedback on the requirements library, from many internal and external partners, takes a lot of time. How do you keep an overview and significantly speed up the improvement process when internal capacity is limited because the critical expertise is scarce? TenneT and predictIT spent the past period attacking this, and this session shows exactly how: mechanism, evidence, worked examples, and the patterns we rejected along the way. We start with an independent requirement review. The INCOSE Guide to Writing Requirements and ISO/IEC/IEEE 29148 already tell us what a good requirement is; the difficulty is applying those criteria consistently, at volume, on time. We decompose each criterion into small, near-deterministic checks, and feed every check the facts it needs rather than trusting a model to recall them. We show how facts retrieved from a linkeddata requirement graph (RDF per NEN 2660, TenneT OTL) is our first step towards governable ai use in the creation of engineering artifacts. In parallel, the feedback coming back from the projects is structured, organised and rated on quality. Combining the two produces an improvement proposal that both meets the standards and reflects what the feedback actually said. A language model is used as glue: sequencing the checks and writing the conclusion. The result is a review that is fast, consistent, and carries its evidence. We close with the lessons learned and the problems we have not solved entirely. Triage scoring, for example, is ordinal, and experienced engineers disagree with each other as much as with the machine, which leaves AI validation without stable ground truth. We will present our current handling and ask the chapter to take it apart.