Rasmus Toivanen
AI CTO
Rasmus Toivanen is the AI CTO at Recordly, bringing nearly a decade of experience in Data Engineering and Machine Learning, with significant expertise in the Insurance and Manufacturing domains (5+ years each).
He specializes in building scalable data and ML systems, generative AI, NLP, and MLOps, focusing on bridging cutting-edge research with practical enterprise applications using cloud platforms like AWS, GCP, and Azure.
His technical expertise spans from data engineering tools (e.g., Databricks, Airflow, Data, SQL, and Data Vault 2.0) to ML technologies (e.g., Tabular modeling, Speech Recognition, NLP, GenAI) .
Rasmus has developed robust data pipelines, implemented Data Vault models and led various ML projects by applying various pre-ChatGPT NLP techniques and modern generative AI solutions to client problems and in addition to that he has won Fennia AI Hackathon (an insurance company).
At Recordly, he drives AI initiatives and works in client projects across various sectors, leveraging his background in Industrial Engineering, combined with his expertise in data and ML. Rasmus offers both proven experience and visionary guidance when it comes to AI.
Finnish-language LLMs
Being a key contributor to Finland's AI ecosystem, Rasmus has been developing multiple Finnish-language LLMs and actively participates in the Finnish-NLP open-source community on Hugging Face.
He contributes to open data initiatives and has experience refining Finnish speech recognition models which have reached over 200,000 downloads to date.
Rasmus often shares insights on the AI and data landscape through articles and LinkedIn posts, advocating for efficient models and emphasizing local innovation. He has also been recognized in the press, such as in Tivi and Mikrobitti for his work.
Rasmus on YouTube: Showcasing how to finetune Finnish language models
Rasmus developed Finnish Ahma language models with 3B and 7B parameters by pretraining and finetuning the models from scratch, based on the Llama architecture.
In this video, he demonstrates how the model can be fine-tuned for specific use cases. The showcased example involves fine-tuning the model to generate plausible headlines from article texts.