I build scalable, robust data platforms, AI solutions and machine learning in production.

I'm Kjetil Dyrland, a data engineer at Tet Digital in Oslo. Before that I moved Lerøy Seafood's data onto Databricks and built models for salmon prices and shipping documents. Through my own company, Dyrland Data & AI, I take on data and AI projects, like invoice control for Entro.

See the apps Read my CV Send an email

Apps

Eight apps on the App Store, seven of my own and one for a client. Built from scratch in Swift. Tap an app to take a look.

Data and AI

My day job: getting data out of old systems, through tested and documented pipelines, and into the models and reports people use. Read the CV

Experience with

Data: Snowflake, dbt, Databricks, PySpark, Delta Lake, Delta Live Tables, Unity Catalog, PostgreSQL, Supabase
Machine learning: Decision theory, Bayesian methods, classifier evaluation, XGBoost, CNNs, MLflow, LLMs
Cloud: AWS, Azure, Cloudflare, Vercel, Docker, GitHub Actions
Languages: Python, SQL, TypeScript, Swift, Rust, Java
Apps and web: SwiftUI, SwiftData, React, Next.js, FastAPI, Xcode Cloud

With Entro

Entro AS helps Norwegian businesses use less energy. I'm building part of the portal their customers use.

Entro:Pi is where Entro's customers follow their buildings, energy use and ESG reporting. I'm building its new invoice control, which replaces an old desktop system.

Every electricity, grid-rent and district-heating invoice is checked line by line against hourly meter readings, spot prices and grid tariffs. When something doesn't add up, it gets flagged. The next phase turns checked invoices into what each tenant owes.

Research

Two papers I co-wrote during my master's, on how to judge and use a classifier when different mistakes cost different amounts.

arXiv, February 2023

Does the evaluation stand up to evaluation? A first-principle approach to the evaluation of classifiers

Kjetil Dyrland, Alexander S. Lundervold and P.G.L. Porta Mana

Accuracy, F1 and MCC can recommend a classifier that loses money. We show that the only consistent way to compare classifiers is to weigh each outcome by what it's worth, and that even a rough guess at those values ranks classifiers better than the usual metrics.

arXiv:2302.12006 PDF

What the usual metrics say

Metric	A	B
Accuracy	0.62	0.75
F1	0.59	0.77
MCC	0.24	0.51
Precision	0.64	0.70

What each one earns per component

Damage when a short-life part ends up in a long-life device335

Confusion matrices and gains from the paper's prologue. Classifier A is right more often about short-life parts, and those are the mistakes that cost the most.

arXiv, February 2023

Don't guess what's true: choose what's optimal. A probability transducer for machine-learning classifiers

Kjetil Dyrland, Alexander S. Lundervold and P.G.L. Porta Mana

A classifier's score is not a probability. We turn scores into real probabilities with a Bayesian model fitted once on held-out data, then pick the action with the best expected outcome. On drug-discovery data this beat the usual 0.5 cut-off in every cost setting we tried.

arXiv:2302.10578 PDF

Standard (score above 0.5)–

With the transducer–

Value of finding an active molecule, compared with the cost of testing a dud10 : 1

Computed in your browser from the paper's published random-forest transducer and 3,589 calibration molecules. The paper's headline numbers come from a separate test set, so they differ a little.