Orchestrating agents without losing the plot
Multi-agent systems fail in the seams, not in the models. What we changed after watching a supervised network quietly drift for three weeks.
2 min read
Nils Aubert
Builds the retrieval and agent infrastructure behind Neuryn's engagements. Ten years in distributed systems before applied AI, which mostly taught him how expensive a bad abstraction gets.
Nils started Neuryn in 2021 after four years building recommendation infrastructure at scale, where he learned that most model problems are actually data-boundary problems wearing a disguise.
He leads architecture on every engagement and writes most of the evaluation code. He is unreasonably attached to the idea that a system you cannot measure is a system you do not have.
Published
2 articles by Nils.
The eval suite is the product
Ship the harness before the feature. Why we now write evaluation code first, and what it costs when you do not.
2 min read