Lumi: a Local-First AI Assistant

CTRLAB.NET - AI, North Macedonia

CTRLAB.NET used Claude, ChatGPT and Google Gemini to design, build and run Lumi, the AI assistant on aibalkans.com, which answers readers from the publication's own content first.

The ChallengeAI Balkans wanted readers to be able to ask a question about AI in the region and get an answer they could trust. Most website assistants send every question to an outside AI, which makes them slow, costly to run and able to say things a publication never wrote. For a news site, one invented answer costs more than the feature earns.The team was also small, with no budget for a large AI bill, and the assistant had to work reliably on free AI services whose limits can change without notice. The AI SolutionCTRLAB.NET used AI to build the assistant, but not to answer every question. Claude supported the design and development of Lumi inside the project's codebase, from the first prototype to every fix that followed. ChatGPT helped with early ideas, including the name Lumi, from the Latin word for light. Google Gemini became one of the services Lumi can call, but only when the website has no answer of its own.Lumi answers from the publication's own content first: articles, categories and tags, the glossary, the tools catalog and hundreds of curated answers. An outside AI is the fallback, not the engine. How It Was ImplementedThe work was done in small stages and checked against the live site each time. The team first built the assistant's own search and curated answers, then added the interface: the source stories shown with every answer, a sheet that opens over the page on phones, and a small character with its own moods, so waiting feels like progress instead of a blank screen.When the free AI services reached their limits and some answers slowed down, the team did not hide the problem. Failures were made visible in the admin, a second provider was added, and finally the AI was taken out of the main path for news questions. Questions about a country were also corrected so that they search articles only, with stories filed under that country shown first. When a country has only two stories, Lumi says so instead of padding the answer.AI suggestions were never treated as finished work. Every change was tested, reviewed and checked on the live site before it was kept. What ChangedQuestions about a country or a topic are now answered in about a third of a second, without calling an outside AI, and the assistant keeps working when an AI service is slow or unavailable. The editorial team can see what readers asked and what Lumi could not answer, and add curated answers from the admin.The biggest change was not speed. It was control: a small civil society organization built an assistant that readers can check, that does not depend on a paid AI service to answer news questions and that only says what the publication has published.