AI tools that work inside the newsroom's own editor
We build AI helpers into the editorial CMS. They draft headlines and intros, write SEO titles, descriptions and image texts, find related articles by meaning and prepare articles for listening. Editors can change every result, and the publisher can switch each tool on or off.
What the tools do
-
Headlines and intros
A button in the article editor offers several variants of a headline and an intro, written from the article itself. The editor picks one, edits it and approves it.
-
SEO titles and descriptions
The tool keeps a headline that already works for search and rewrites or shortens one that does not. It writes the description from the article's intro and gives a reason for each decision.
-
Image texts
An ALT text and a title for the article's main image, written from what the article is about. The caption the editor wrote stays as it is.
-
Related articles
Three articles to read next, chosen by meaning from the whole archive and mixed with what is recent and what is being read.
-
Articles to listen to
The text of an article is prepared for speech synthesis and played in an audio player on the portal.
Work
-
Ir.lv
AI tools in the newsroom of a weekly magazine, covering headlines and intros, SEO texts, related articles and audio.
Read the case study: Ir.lv -
Lasi.lv
Headline and intro variants, SEO suggestions, related articles and narration in the newsroom of a national publisher.
Read the case study: Lasi.lv -
E-klase
An education news portal for teachers, parents and students, with an editorial CMS and site search.
Read the case study: E-klase
From a print headline to a search headline
A magazine headline is often a metaphor that works on a cover and tells a search engine very little. The SEO tool reads the headline, the intro, the section and the beginning of the article and decides whether the headline can stay.
If it is clear and names what the article is about, the tool reuses it. If not, it rewrites it in a neutral news style, with the people, places and organisations the article mentions. It adds nothing that is not in the article.
The description shown under the title in search results is made the same way from the article's intro. Before this, descriptions were cut from the first sentences and often ended mid-thought.
-
Reuse before rewriting
A text that is already good is left alone. The tool changes only what does not work.
-
A reason for every change
Each result comes with a short explanation, so the editor can see why a headline was replaced.
-
Only when something changes
Texts are generated again only when the article's headline or intro changes.
Related articles by meaning
Tags and keywords miss articles that cover the same subject in other words. An article about the rising cost of a railway project and an analysis of tax reform share a theme and hardly any words.
We turn each article's headline, intro and text into a vector, a numeric fingerprint of its meaning, and store it in PostgreSQL with pgvector. Articles with similar fingerprints are close in meaning.
Closeness alone would keep suggesting old material, so the three places under an article are filled by different rules: one favours the closest match, one balances closeness with how recent the article is, and one brings in what readers are opening now.
- Article saved
- Text to vector
- Vector search
- Closeness with recency and popularity
- Three suggestions
-
The whole archive
Older articles that are still worth reading are found again, without an editor searching for them.
-
No manual linking
Suggestions are prepared when an article is saved. Editors do not pick related articles by hand.
-
Sensible fallbacks
The open article is never suggested, nothing is shown twice, and a place with no good candidate is filled by another rule.
Preparing content for speech synthesis
Speech synthesis reads exactly what it is given. Numbers, dates, abbreviations and names come out wrong unless the text is prepared first.
For Latvian we write numbers and dates out in words, expand abbreviations, correct the pronunciation of proper names and add pauses where a listener needs them. The result is generated in SSML, the standard format for speech.
Editors generate the audio in the CMS and choose what is offered for listening. The portal's player lets the listener change speed, jump within the recording and set the volume, and the files are delivered through a CDN.
One layer between the CMS and the models
The tools do not call language models directly. They go through a service layer of ours, so the publisher is not tied to one provider.
-
Providers can be changed
The layer gives the CMS one interface to several model providers. A provider can be switched in the settings, and a backup takes over if the main one fails.
-
Every call is logged
The log shows what was asked, what came back, how long it took and what it cost.
-
Limits on use and cost
Request and cost limits are set for the whole system and for each model.
-
A switch for each tool
An administrator turns each AI tool on or off in the CMS settings. Texts generated earlier are kept.
-
The CMS works without the AI tools
If the LLM models are unavailable, editors carry on and publish without the AI helpers.
-
Data security
Only the text a task needs is sent to the models, and it is cleaned before it leaves the CMS. Everything else stays in the publisher's system.
AI for other kinds of work
The same building blocks apply outside newsrooms. Generative AI inside an existing product, search over a company's own documents and advice on where AI is worth using are described on the service page.
Want AI helpers in your content workflow?
Tell us which routine task takes your editors the most time.