About
I'm an AI/ML Engineer and Data Scientist completing an MSc in Artificial Intelligence in Oslo. I build machine learning systems and put them in front of people: every project here runs as a live demo, and one is a production website for a real client. My focus is explainability: making AI decisions understandable and trustworthy, across anomaly detection, predictive modelling, time-series, and computer vision.
What I've Built
-
Retrieval-augmented question answering over 977 pages of Equinor, DNB and Hydro annual reports. Every answer cites its page, and the system refuses with an explanation when the reports do not contain the answer. Scored 93% correctness on 20 unseen, human-verified questions.
-
Churn drivers, a value x risk targeting grid and a retention test plan on 7,043 telecom customers, plus a real 64,000-customer A/B test where uplift targeting gave 8.4 points of lift against 3.4 for a "most likely to respond" list.
-
Unsupervised segmentation of 678,013 French auto insurance policies using K-Means and PCA to surface three actuarially distinct risk profiles, validated with Kruskal-Wallis.
-
Deep learning pipeline (1D CNN, LSTM, and a CNN-LSTM hybrid) to predict forward realized volatility of SPY over 5- and 10-day horizons using 33 years of price history.
-
PySpark pipeline over 2020-2024 NYC Yellow and Green Taxi trip records: Parquet ingest, Spark SQL cleaning, feature engineering, and a Spark ML regression to predict fare amount. Companion Gradio app replays the EDA queries against persisted SQLite samples.
-
Four studies across data mining and computer vision: network graph mining on movie data, NLP classification on farm advertisements, low-light image enhancement, and AR landmark recognition.
-
End-to-end supervised ML pipeline for heart disease prediction: preprocessing, logistic regression, random forest, 5-fold cross-validation, and unsupervised pattern discovery.
-
Hybrid anomaly and signature-based IDS on the CIC-IDS-2017 dataset, with SHAP and LIME to make model decisions interpretable for security analysts.
-
Logistic-regression churn model on a 10 000-row European retail-banking dataset, wrapped in a live what-if calculator: pick a real customer, nudge their attributes, watch the prediction and driver breakdown update in real time.
Selected Builds
-
A full-stack marketing and booking website for a California hair studio, live in production. Built with Next.js, Supabase, and Tailwind and deployed on Netlify, with a services and pricing system driven from a single source of truth, an image gallery, email signup, and an integrated booking handoff.
Experiments
-
A Norwegian language-learning tool built through iterative LLM collaboration: type any word or phrase and get a full linguistic entry (translation, IPA, all inflected forms, examples, grammar notes, and slang detection), then save what you're weak at and track high-frequency-word progress.