At K-Resourcing - Haime, I replaced the prompt classifier with a multilingual transformer model, reducing median latency from 15 seconds to 241 milliseconds and nearly halving memory use.
I built an automated classifier evaluation pipeline with a hand-labelled test set, language-level accuracy checks, and error analysis. It helped me stop a model change that performed worse than the production model.
I developed Python services to detect personal data and secrets in English and Danish, and deployed and maintained ML services on Kubernetes. I also diagnosed production incidents and developed features for the main TypeScript API, including LLM integrations and document search.
In my projects, I built a multimodal scientific-paper RAG system and an SMS phishing detector. My work has included model evaluation, retrieval pipelines, and containerized services for production use.
