Sara Moricz, PhD · Research Data Scientist · Skåne, Sweden

Sara Moricz

I work on machine learning close to the data: sensor time series, anomaly detection, and models that run privately on the device. Right now I'm writing about small language models and why they struggle with Swedish.

Subscribe on Substack LinkedIn

Now updated 10 Nov 2026

Ordverkstaden (in Swedish) is open: chat with a small model steered to invent words.
Starting a series on small language models and Swedish. First piece: the token tax, how much more it costs to use AI in your own language.
Getting the Android beta of WithYou ready for its first testers.
Research project: federated learning for language models, so models can learn from local text without the text leaving the device.

Projects

Two dictionary cards the model wrote for words it invented. Hjärtklyftande: something so emotionally strong it cuts through the heart. Skåpfull: a playful twist of the word skåpsfulla. Each with synonyms and an example sentence.

Ordverkstaden

Chat in Swedish · Nov 2026

The upside of the token tax: a model that builds Swedish from pieces learns to build new words. Most are bad, some brilliant. I found where Qwen 3.5-4B coins new words and turned it up. Chat with it: can a small model be creative? (In Swedish)

Chat with it →

Three glowing dot globes coloured by token tax: English-speaking countries in teal, most of Europe and Asia in violet and pink, Ethiopia and parts of Southeast Asia in orange and yellow.

The token tax

Interactive globe · Oct 2026

Every language pays more than English to use AI. Swedish needs 35 % more tokens for the same text on ChatGPT, Hindi 57 %, Amharic over 480 %. Measured on 204 languages and 13 tokenizers, with open code and data.

Spin the globe →

WithYou

Android · beta coming

A safety app for older people and their families. It learns what a normal day looks like on the phone and alerts a relative when something seems off. Anomaly detection runs on the device and no location history is stored.

Background

NowApplied ML research: time series, anomaly detection, federated learning, differential privacyHorizon Europe and Vinnova projects
2019–2022Data scientist, IKEA DigitalCausal inference and a global decision-support tool
2013–2019PhD in Economics, Lund UniversityIncome inequality and policy evaluation; one chapter used NLP to track gendered language on Swedish Twitter after #metoo

New pieces by email

For now I write about small language models, Swedish and machine learning close to the data. I'm a curious generalist: economics, statistics and AI in one toolbox, pointed at what technology means for people. Concise and illustrated.

Subscribe on Substack