Capability
Content pipeline
Everything a learner reads, watches or practises passes through a pipeline first. Articles are collected, read by an AI, turned into words and grammar to learn, and published. Videos are made in a studio, transcribed and captioned. Practice is generated and checked before anyone sees it.
Articles in, learning material out
Collect
A scheduled ingest job loads news and article sources, extracts the article text, runs an AI quality check and files drafts in the content management system. Editors control each source: on or off, how many articles per run, and which run slots it takes part in. High-confidence article pages are read by a deterministic extractor first; the AI extracts only when that is not enough.
Understand
When a draft is saved, the ingest processor reads it in two passes. The first picks the words and grammar concepts worth learning from that text, with prompts written per language. The second finds every occurrence of each pick in the text and writes an explanation for that occurrence.
Publish
Results return to the content system as events. Editors review and publish. Learners then read the article at their level, tap any word, and get its meaning, forms and the explanation written for that spot.
Video and audio, transcribed
Speech in a video is turned into text with Whisper, running on Kielo's own infrastructure. The language of the transcript comes from what the model hears, not from what was requested, so a wrong label cannot pass as a correct transcript. Heavy transcription is rate-limited across instances, backs off when the transcriber is busy, and hands a message back for redelivery instead of marking a video as failed.
Transcripts become captions: two lines per cue, in the language being learned and in the learner's own language, with tappable words. A carousel has no audio, so its hand-written captions stand in for the transcript and go through the same word and grammar analysis.
Lessons and exercises, generated and checked
The learning engine generates lessons and exercises with AI models. A generated exercise is not served on trust: it passes through a fixed pipeline first.
Deterministic gates
Free checks on structure, a present answer, valid options, grammar guards (for example Finnish negation agreement), and a prompt that does not give away the answer.
Answer-key judge
A model checks that the keyed answer is right and that no other option is also right.
Publish, hold or reject
Confident passes are published. Defects are rejected. Anything uncertain, and any judge error, is held and never served until a person approves or rejects it in the admin app.
Related words and grammar are grouped into concepts, and each concept can have a hub: a page with a lesson and a practice deck. Practice drawn from decks reuses published exercises, so it does not call an AI model at the moment a learner taps.
KieloTV: short video that teaches
KieloTV is a vertical short-video feed inside the app, with tabs for daily words, conversations and carousels. Captions are bilingual and tappable, a vocabulary panel lists the words and grammar of the clip, and a learner can save a clip or report a problem.
Editors make a video in KTV Studio, in the admin app. A rendering toolbox produces it, media processing makes the renditions phones need, transcription and analysis run, and the content service serves the feed.
Social posting, scheduled from the admin
Short clips and carousels also go out on Instagram, YouTube and TikTok, for the Finnish and Swedish channels. Editors plan posts on a timeline and month calendar in the admin app. Kielo itself holds the schedule and hands each post to the publishing tool Buffer 48 hours before it is due; Buffer publishes. Exporting a word video drafts a post per connected channel for that language, and failures raise an alert to the team.
Media delivery
Uploads go through one media pipeline. A processor builds the variants each surface needs, such as resized images, thumbnails and video renditions, and each media type has a profile that sets its allowed formats, sizes and access rules. Phones are served an adaptive-bitrate stream where available and a mobile rendition otherwise, rather than the full-size original.