Machine Learning Engineering • Based in Norway • Available Worldwide
I build custom machine learning systems: computer vision, time-series, deep learning, and edge AI. My focus is closing the gap between a promising demo and a system your users, clients, and regulators can actually rely on.
Hi, I'm Kim.
A freelance machine learning engineer based in Oslo, building custom models for high-stakes domains.
More about me ↓Sound familiar?
Sensor streams, images, operational logs, years of it, but no clear path from raw data to a reliable model or scoring framework. You need someone who can turn that data into something that actually predicts, detects, or scores what matters.
The model looked great in the notebook. In production it meets messy, drifting, real-world data it never saw in training, and accuracy quietly falls apart. Closing that gap between benchmark and reality is where most projects stall.
Investors, regulators, or a product gate need evidence that your data actually predicts what you claim. That calls for rigorous modeling and an honest evaluation: real metrics, sound validation, and a clear read on whether the signal is strong enough to bet on.
A notebook that runs once is not a system. You need a robust pipeline, an honest evaluation strategy, and a realistic path to deployment, plus a straight answer on what's worth building. I'd rather tell you a problem isn't an AI problem than sell you a model that won't hold up.
How I solve it
You don't have a generic AI problem, you have a specific one. Forecasting demand, sales, or energy load. Spotting anomalies in sensor data, or detecting objects and defects on a production line. Predicting equipment failure before it causes downtime. In clinical settings, modeling biomarkers and patient risk. I build custom models for problems like these, using computer vision, time-series, and deep learning, and engineer them to run reliably in production, or on-device at the edge when latency, cost, or privacy demands it.
Your model performs well on paper, but stakeholders don't trust it and regulators can't audit it. Explainability is where my PhD work lives. Whether it's medical image classification, clinical risk scoring, or any decision that has to stand up to review, I build models that show their reasoning, using proven techniques like Grad-CAM, so every prediction is auditable and meets both clinical standards and EU AI Act requirements. Proven in medical diagnosis, applicable anywhere a decision carries weight.
You need a knowledge assistant over your internal docs, a document Q&A system, or contract extraction you can trust, but off-the-shelf demos hallucinate, loop, and quietly run up costs. I build production-grade RAG and multi-agent systems with confidence scoring and real evaluation, designed to be reliable, fast, and measurable. Not just impressive in a demo.
Not ready for a full build? Start with a clear-eyed assessment. I help you decide whether AI fits your problem, which approach is worth the investment, and what a realistic path to production looks like. No vague strategy decks, just concrete technical direction you can act on.
Selected work
Object detection on dense, cluttered scenes, the kind of problem retail and logistics teams face when they need to count, locate, and identify products at scale. Trained with YOLOv8 at 1280px to hold accuracy on small, tightly-packed objects, and validated at Norway's national AI championship (NM i AI).
Doctoral research on infant movement recordings, where the input is a pose sequence over time rather than a single image. Spatio-temporal graph convolutional networks identified limb velocity as a candidate early biomarker for cerebral palsy, paired with Grad-CAM, CAM, and perturbation analysis so clinicians can see what the model based its decision on. Developed and validated with St. Olavs Hospital.
Before a wearable or clinical model can be trusted, the signal feeding it has to be usable. A one-metric screening check for four physiological signals: kurtosis for ECG, autocorrelation periodicity for PPG, signal-to-noise ratio for EMG, and paired amplitude and kurtosis for EEG. Each metric is validated against public databases carrying expert labels or a known noise schedule, so the method is measured against an answer key rather than asserted.
A RAG pipeline and a LangGraph research agent over the EU AI Act, built so every statement points to its source. The Q&A pipeline cites the exact article, checks each citation against the text it retrieved, and declines when the Act does not answer. The research agent scores each claim from its sources, and every report goes through a reviewer tested with planted errors.
Tech stack
Writing
September 28, 2026
A step-by-step walkthrough of a question-answering system over the EU AI Act: scoping it, writing the test first, measuring retrieval, checking every answer, and changing one thing at a time, with the results of each step.
September 28, 2026
A step-by-step walkthrough of a research agent that writes cited reports on the EU AI Act: choosing how complex it needs to be, giving it the right sources, scoring claims in code, adding a reviewer, and testing that reviewer with planted errors.
About
Signal Syntax is my machine learning consultancy. I build custom models for teams whose problems don't fit off-the-shelf tools: computer vision, time-series, deep learning, and edge AI.
I hold a PhD in Explainable AI from NTNU, where I built and validated medical diagnosis models that had to be both accurate and interpretable. That work is the backbone of how I approach every project: rigorous modeling, explainable outputs, and results you can defend.
Before research, I spent five years as a hardware engineer designing and shipping production electronics, where "production-ready" was an engineering claim, not a marketing one. That's why I think about models from the ground up: the data, the deployment target, the edge cases, and the people who depend on the output.
Based in Norway. Available worldwide.
Get in touch
Whether you need a quick feasibility check or a full build-out, I'd love to hear from you. The fastest path is a free 30-minute discovery call. Pick a time that works for you.
Book a free call →Prefer to send a message? Use the form, or email me directly at click to reveal address. I typically respond within 24 hours.