AI Engineer in Kristianstad, Sweden. I build agent pipelines and machine learning systems in Python, usually from scratch: hand-rolled async orchestration in preference to a framework, unless the framework solves a hard problem better than I can (see Prasine Index, where LangGraph earns its place).
Finishing a Data Science degree at EC Utbildning in May 2026. Open to roles.
Predicts which Swedish party wrote a parliamentary speech. Fine-tuned KB-BERT at 0.628 accuracy and 0.619 macro-F1, evaluated on 146 speakers held out of training, served from FastAPI on Cloud Run over a 75,148-speech Riksdag corpus that rebuilds from source. The site carries a study that turns the classifier into an instrument: which party do fourteen frontier language models write like?
Prasine Index · prasineindex.com
Seven-agent pipeline that checks EU corporate sustainability claims against 22 open datasets and returns a scored evidence chain. Anthropic SDK and LangGraph.
Async AI orchestration for interactive storytelling, built without LangChain. Deterministic game logic decides what happens; the LLM only narrates it. Voice benchmark: 2.0s TTFA p50 over 46 runs, too slow for conversational turn-taking. 182 tests.
Near-miss prediction for road construction zones. Random Forest at F1 0.927, benchmarked against XGBoost, LSTM and TCN. Numbered pipeline, 01 to 06.
- Hugging Face: models and datasets
- cm.blomqvist@gmail.com