STELLAR

When a single model call answers the customer directly, its answers are difficult to explain, a hallucination goes straight to the customer, and there is no structural place to check compliance or to decide when a person should take over.

STELLAR (Structured, Trustworthy, and Explainable LLM-Led Architecture for Reliable Customer Support) replaces the single call with a directed acyclic graph of nine specialized modules, composed into eleven predefined workflows. Together the modules cover:

• Few-shot classification
• Retrieval-augmented generation (RAG)
• Sentiment analysis
• Urgency-aware human escalation
• Compliance verification
• User interaction validation
• Semi-automated knowledge base refinement

Every request starts at classification and takes one of three branches: retrieval-augmented generation, a direct-information answer, or sentiment analysis. The first two pass through compliance and verification modules, and a failure at either point sends the conversation to a person or to an FAQ path. The sentiment branch leads to human escalation, which feeds knowledge base refinement.

I designed, developed and led STELLAR at Semantix AI as a production-ready architecture, then published it as first author with Hélio Pedrini in the Journal of the Brazilian Computer Society, vol. 32(1), pp. 128-144, 2026. The journal is Qualis A2 and indexed in Scopus.

Separately from STELLAR, I built a hallucination benchmark that evaluated 7 models across 90,000+ questions in English and Portuguese. I also researched RAG techniques and built data pipelines for RAG-based retrieval.