View demo ↗ The Challenge
Dozens of repetitive queries about academic statutes, procedures, and bylaws arrived daily via WhatsApp and forums without the capacity to answer in real time.
What I did
A Qwen 3.5 (4B) fine-tune on UNAL regulations with a RAG engine over official sources, a dedicated API with streaming, and multi-channel deployment: WhatsApp, CEIS chat and forum. Every answer cites its sources.
Result
Live chat at repre.nikko.dev and a production API, answering regulations with cited sources across all three CEIS channels.
What I learned
In production RAG systems, semantic chunking and source provenance matter far more for precision than raw model parameters.