- Model: Fin Apex 1.0, post-trained on proprietary production support data, replacing foundation models
- Vs Claude Sonnet 4.6 in production: +2.8% resolution rate, 0.6s faster latency, 65% fewer hallucinations
- Pipeline: 7 discrete models: language detection, issue summarization, knowledge retrieval, result reranking, response generation, feedback parsing, escalation routing
- Retrieval: RAG based, retrieves ~40 candidate documents, reranks, filters context before generating
- Hallucination mitigation: grounded in retrieved context (not pretraining knowledge), a dedicated hallucination checker at generation time, a final AI Engine safety/accuracy check
- Published research: actor-critic approaches to reduce hallucinations, structured agentic RAG for ecommerce, topic modeling for dialogue