AI content processing and recommendations for the Kyiv Independent
For the Kyiv Independent, we automated article summaries, tagging and related-article recommendations with LLMs and embeddings — improving the site’s UX and search visibility while cutting processing costs.
- Industry
- Media, Publishing
- Region
- Ukraine
- Services
- Data science, LLM engineering, Recommendation systems
- Duration
- 3 weeks
- Project duration
- 3 weeks
- Senior data scientist
- 1
- Summary and tags generated together
- 1 call
Intro
The Kyiv Independent is a leading English-language media outlet based in Ukraine, publishing independent, fact-based journalism on Ukrainian and global affairs.
We worked with The Fourth Estate — our direct client, which provides the Kyiv Independent’s technology platform — to optimize summarization, tagging and recommendations with AI models, improving processing efficiency while reducing costs.
Goals and challenges
- 01
Editorial alignment
Output had to follow the Kyiv Independent’s editorial standards and guidelines.
- 02
Complex tagging
Tags had to work across several formats, from free-form to a fixed schema.
- 03
Consistent tone
Summaries had to match the editorial voice consistently, which took iterative testing and refinement.
Solution
We tested several approaches for summary and tag generation, then combined the best of them into one pipeline.
Summary generation
Arbitrary summaries come from an LLM tuned to the editors’ style; newsletter summaries use few-shot learning to match the newsletter voice.
Tag generation
Four modes: free-form tags, tags guided by editorial policy, static schema-only tags, and a hybrid of schema and policy.
Cost reduction
Summaries and tags are generated in a single model call, which improves performance and reduces API usage.
Recommendations
Content-based recommendations use embeddings in a vector database, with similarity scoring and category filtering.
Team
- 1 Senior Data Scientist
Results
- Content processing is automated, more efficient and cheaper, and integrates with the existing system.
- AI-driven, guideline-based or predefined tags improve site search and SEO.
- Free-form and newsletter-style summaries improve readability and AI-detection rates.
- Embedding-based similarity delivers relevant article suggestions.
- The team was responsive, completed all tasks on time and communicated through virtual meetings, email and messaging apps.
Project details
- Client
- The Fourth Estate, technology platform for the Kyiv Independent
- Industry
- Media, Publishing
- Region
- Ukraine
- Services
- Data science, LLM engineering, Recommendation systems
- Technologies
- Python
- SQL
- OpenAI
- LangChain
- Docker
- Vector database
- Team
- 1 senior data scientist
- Duration
- 3 weeks
