Enterprise Knowledge Assistant · Conversational RAG (HPE)

I led an enterprise conversational AI assistant for support engineers across HPE. It retrieves grounded answers from roughly 1,700 internal process documents sourced from a Squidex content management system and Confluence, and cut repeat escalations by ~40%. Over the project it moved from rule-based FAQ retrieval to an LLM-assisted, retrieval-augmented architecture, and shipped as a dockerized deployment behind a load balancer, sized for a 20,000-concurrent-user target.

Support engineers burned time hunting for process documentation, and the rule-based FAQ system fell over on anything outside its predefined set. Escalations repeated because the answers were hard to find, not because they were missing.

I built the case for the redesign off customer usage analysis, architected the migration from rule-based FAQ to LLM-assisted RAG, designed the retrieval pipeline end to end, and owned the rollout.

1. Hybrid search over pure vector retrieval. Enterprise-jargon queries break cosine similarity; combining semantic and keyword signals with metadata filtering rescued them.
2. Chunking with overlap calibration. Retrieval quality was tuned at the chunk boundary level, not just the embedding level.
3. Re-ranking on top of hybrid recall. One extra hop of latency, traded for a 30-40% response relevance lift measured in pilot.
4. Evaluation loops on a held-out set. Retrieval precision was measured against a fixed evaluation set on every pipeline change, so improvements were provable rather than anecdotal.
5. Incremental migration, not a rewrite. Evolving the live rule-based system kept the escalation baseline measurable and de-risked the cutover.

Repeat escalations down ~40%. Response relevance up 30-40% in pilot against a held-out evaluation set. Deployed as dockerized services behind a load balancer with continuous ingestion from Squidex and Confluence.