Applied AI · Gothenburg, Sweden
We build focused consumer products — for studying, job hunting and everyday life — on top of speech and vision models we finetune ourselves. We put our energy where the large labs don’t: smaller languages, local needs, and unglamorous problems worth solving properly.
2025 — Gothenburg, Sweden
Bootstrapped — products live in market
Swedish & low-resource-language AI — speech, vision, everyday tools
Each one started as a problem we kept running into ourselves.
minnefy.se
Turns a YouTube link, a PDF, a photo of your notes or an audio file into flashcards and quizzes — then makes sure you actually remember them.
talanzo.se
A career platform built for the Swedish job market: an agent reads the job boards every day, scores openings against your CV, and drafts the letter.
Swedish dictation
Speech-to-text that was tuned for Swedish rather than translated into it — built on open, commercially usable models we finetune in house.
Grocery price comparison
What the same basket costs across the major Swedish grocery chains, indexed nationally and kept current — with a mobile app in progress.
Why we exist
Swedish has around ten million speakers. Dari has more than thirty-five million. Both get whatever quality happens to fall out of a model trained for somewhere else. Open models plus targeted finetuning now beat the generic APIs on exactly those languages — at a fraction of the cost. We ship consumer products on top of that gap and keep the models we train.
Speech and vision models trained in house, on consumer and rented NVIDIA GPUs.
Real Gothenburg driving-exam routes reconstructed as 3D Gaussian splats from car-mounted video, and drivable in the browser. Splats carry the visuals; an engine layer carries physics and traffic rules — the same hybrid split the large simulation stacks use.
Whisper-family models finetuned for Swedish on top of KB-Whisper — an openly licensed, Swedish-optimised model that already outperforms the general-purpose flagship on this language. The output goes straight into dictation and into study tools that have to hear Swedish properly.
A first speech benchmark and model track for Dari — spoken by more than 35 million people, with effectively no dedicated datasets or models today. Built from the ground floor by a native speaker, and released openly.
Questions go straight to the founders — usually answered the same day.
A question about a product, an idea worth building, or a partnership — we’d like to hear about it.
Prefer email?
hello@rezify.se