This page is just a backdrop. The chat widget in the corner is the artifact: a production-grade Agentic RAG bot with hybrid retrieval, brain intent routing, and zero-hallucination structured tools. Click it to talk.
WhatsApp channel runs the same backend orchestrator and knowledge base; live WA demo available by request.
What the bot actually does
Single brain LLM call decides: refuse, handoff, tool call, or RAG. Skips retrieval 60% of the time.
BM25 + Qwen3 dense embeddings + RRF fusion + LLM-as-reranker. 81% accuracy on complex queries.
Rate limit, regex injection guard (12 EN/ID patterns), PII output masking, per-session token budget.
Production stack