AI, machine learning and LLM systems. MSc in Bioinformatics and Applied Statistics, Oslo.

Projects
Selected work

Projects

Three things I have built: a library from my thesis, a system that reviews reports with LLMs, and a public data service for AI assistants.

01 · Open-source library · From my thesis

expert-priors

Can an LLM’s domain knowledge help a model survive distribution shift?

Ordinary regularisation shrinks every coefficient toward zero. expert-priors asks a language model what the coefficients should look like, turns its answers into a prior, and shrinks toward that instead. Cross-validation decides how much to trust it, so when the data speaks clearly, the model follows the data.

4 / 5datasets improved out of distribution
7.9%improvement out of distribution, best case
3prior families, all scikit-learn estimators
  • Python
  • scikit-learn
  • Claude
  • OpenAI
  • Gemini
Regularisation paths for four coefficients. Plain Ridge shrinks each one to zero as the penalty grows, while the prior-informed model shrinks toward the LLM's prior instead.
Ridge shrinks to zero; the prior-informed model shrinks to the LLM prior.
import expert_priors as ep

# Ask the LLM for coefficient draws (cached)
priors = ep.elicit_priors(X, y, task="regression",
                          auto_context=True)

# Fit like any scikit-learn model
model = ep.PriorMeanRidge(priors).fit(X_train, y_train)
02 · LLM evaluation system

AI Report Evaluation

Can an LLM review a building inspection report the way an expert would?

Quality control for Norwegian building inspection reports. Deterministic rules, an LLM judge and retrieval from previously reviewed reports grade each one as clean, minor error or major error. When rules and model disagree, the report is flagged for a person to decide.

In agentic review mode the model looks up rules and examples, reruns checks and verifies passages before it answers. Reports can also be drafted from site visits, with voice notes and photos.

3grades: clean, minor, major
Agenticreview that checks its own work
Voiceand photos to a draft report
  • Python
  • Flask
  • SQLite
  • OpenAI
  • Anthropic
  • Pydantic
Report #8
A reviewed report: the rule-based check says major error, the LLM says clean, and the report is flagged for a reviewer to accept or override.
03 · MCP server and JSON API

Kildeverk

What if every AI answer about Norway came with its source?

Norwegian public data, straight from the source. A free service, with no account needed, that gives AI assistants and developers current figures from SSB, the Meteorological Institute, NVE, Kartverket, Entur, Brønnøysundregistrene and more. Every answer states where its data came from.

30data sources
17publishers
62tools
  • Python
  • MCP
  • Pydantic
  • SQLite
  • Docker
  • Fly.io
Research

Papers

Preprint · arXiv · 2026

LLM knowledge for improved Out-of-Distribution Tabular Prediction

With Kristoffer Herland Hellton and Hugo L. Hammer. LLM-generated priors as the shrinkage target in regularised regression improved out-of-distribution prediction on four of five datasets.

Master’s thesis · NMBU · 2026

Using LLM Knowledge for Improved Out-of-Distribution Tabular Prediction

Graded A, written in collaboration with Simula Research Laboratory. Three estimators that bring LLM-elicited priors into regularised linear and logistic regression.

Building something with agents or LLMs?

I would love to hear about it.

© 2026 Sølve Langseth · Oslo