Gosta Labs Oy is developing the Gosta Labs Oy AI Operating System: an AI companion built with healthcare and social care professionals to reduce documentation burden, structure fragmented clinical information, and support safer, higher-quality care workflows. Gosta Labs Oy’s main product is an AI medical scribe that uses clinical automatic speech recognition to convert patient-clinician conversations into text, summarizes them into clinical documentation, and applies natural language processing models for classification, extraction, and structuring of relevant medical information.
A major pain point for Gosta Labs is the current reliance on third-party proprietary AI models, including foundation models and speech recognition systems. This creates significant challenges regarding data sovereignty, regulatory control, transparency, cost efficiency, and the ability to adapt models to specific clinical workflows, languages, and documentation requirements, especially in non-English European clinical environments.
Previous work on these models has already reduced dependency on commercial providers, and the central objective of this project is to reduce this dependency further. This will be done by developing an in-house clinical AI model stack and continuing the work of replacing key parts of the current workflow with locally controlled models that can be trained, evaluated, and governed within a regulated deployment framework.
The requested compute will be used to train, adapt, and evaluate three connected categories of models: clinical ASR models for healthcare and social care conversations; clinical language models for summarization and documentation, such as patient notes, historical summaries, and referrals; and specialized classification and information extraction models for structuring clinical concepts such as diagnoses, medications, allergies, procedures, and measurements.
The work performed on LUMI will use synthetic data and non-personal derived artifacts only. No real patient data, identifiable data, pseudonymized data, or sensitive health data will be transferred to or processed on LUMI. Any real-world clinical material used outside LUMI will remain within Gosta Labs' secure environment and will not be part of the requested LUMI allocation.
A key technical motivation is improving model quality in multiple European languages, including low-resource languages such as Finnish and Estonian, and in domains such as oncology, which are poorly served by general-purpose proprietary models. Clinical language is highly domain-specific, involving shorthand, specialty terminology, abbreviations, and noisy transcripts.
Developing these in-house models also supports regulatory objectives, aligning AI development with frameworks such as the EU AI Act and ISO 42001, strengthening the ability to control data flows, validate model behavior, and maintain auditability.
The expected result is a more sovereign, clinically adapted, and deployable European healthcare AI stack for speech recognition, documentation, summarization, classification, and information extraction. This will reduce administrative burden, improve the structuring of fragmented clinical information, and form the basis for future human-supervised clinical workflow support.