Data platforms, machine learning, and the products built on top of them.
Dyrland Data & AI is a software company in Bergen. I build production data infrastructure and machine-learning pipelines for Norwegian industry — migrations off on-premises systems, streaming pipelines, models taken all the way into production — and I ship consumer apps on the same stack.
One person, end to end: the Swift client, the API, the database, the deployment.
- Live on the App Store
- 5 apps
- Peer-reviewed papers
- 2
- Registered
- Jan 2026
- Based in
- Bergen, NO
Everything shipped, and everything currently in build.
Filled marker means it is live and you can download it today. Hollow means it is still being built.
Keeps the peace in families who share a cabin. A fair calendar rotates the popular holiday weeks between owners, shared costs are split and settled through Vipps, and open/close checklists keep the place safe. Works offline. Swift · SwiftUI · Next.js · Cloudflare D1 & R2
App Store → TalearkLive · iOSNorwegian voice-to-document for tradespeople. Two minutes of speech after a site visit becomes both a structured inspection report and a priced quote — hourly rate, materials, VAT, declaration of conformity — as a branded PDF sent before you reach the van. Swift · SwiftUI · Next.js · Cloudflare D1 · speech-to-text & LLMs
App Store → GlimtLive · iOSA disposable camera for events. The host hangs a QR poster and every guest gets a 25-shot roll in their phone browser without installing anything. Photos develop with a retro look and land in a shared darkroom the host releases when the night is over. Swift · SwiftUI · Next.js · Cloudflare D1 & R2
App Store → UbruttLive · iOSA screen-time coach that never blocks. It puts a deliberate pause in front of the apps you chose, shows your own stated reason and today's count, then lets you decide — keeping an honest ledger of how often you opened and how often you turned back. Swift · SwiftUI · SwiftData · App Intents · Cloudflare D1
App Store → MosvoldLive · iOS + PWA · client workProperty management and tenant services, built for Mosvold & Co. AS. Role-based login, property information, messaging with the property manager, a benefits programme. Norwegian and English, native on iOS and an installable PWA on Android. Swift · SwiftUI · Next.js · Supabase
App Store →A copilot for the ninety seconds before and after a meeting, built for ADHD work days. Transition alarms at your own offsets, commitment capture when a meeting ends, placement into real calendar gaps, honest rescheduling. Local-first — nothing leaves the phone. Swift · SwiftUI · SwiftData
clearhour.appA social-deduction party game of secret number patterns, bluffs and accusations. One player knows the rule, one lies about it, and the board always carries exactly two lies. Pass-and-play on one device, or online from everyone's own phone. React · TypeScript · Supabase realtime
zebragame.appA real-time value-bet and arbitrage finder. Uses a sharp bookmaker as the baseline, compares odds across Norwegian and international books, and surfaces positive expected-value opportunities with Kelly-criterion stake sizing across 54+ leagues. React · Python · Rust · Supabase
tippegutta.no →Three kinds of work, one stack underneath.
Getting the data somewhere useful
Migrations off on-premises infrastructure onto Databricks and Azure. Pipelines in Delta Live Tables and PySpark, governance through Unity Catalog, and the unglamorous work of making a platform something a team will actually trust.
Models that reach production
Price forecasting, OCR for document processing, retrieval-backed assistants. The research background is in how classifiers should be evaluated, which mostly shows up as a refusal to ship a model on a metric that does not match the decision it feeds.
Apps, end to end
Native iOS in Swift and SwiftUI, with edge backends on Cloudflare Workers, D1 and R2 or on Supabase. Five apps in the App Store, each taken from an empty repository to review by one person.
On evaluating classifiers properly.
Two papers with Alexander S. Lundervold and P.G.L. Porta Mana, arguing that the standard way of scoring machine-learning models is not merely imperfect but avoidably wrong.
Does the evaluation stand up to evaluation? A first-principle approach to the evaluation of classifiers
Shows that popular metrics — F1-score, MCC, AUC — are mathematically never optimal: each produces an avoidable fraction of incorrect evaluations. Makes the case for grounding evaluation in decision theory with problem-specific utilities.
Don't guess what's true: choose what's optimal. A probability transducer for machine-learning classifiers
Proposes calculating probabilities from a trained classifier's output rather than from features directly — a one-time "transducer" enabling decision-theoretically optimal choices. Demonstrated on a drug-discovery problem with heavily imbalanced data.
Available for data platform, ML and product work.
Kjetil Dyrland — data engineer and ML engineer. M.Sc. in Software Engineering and Machine Learning (HVL / University of Bergen), Databricks Certified Data Engineer Associate. Previously data engineering at Lerøy Seafood and consulting at Bouvet ASA.