> forward_deployed
Building the Production RAG Pipeline
Demo RAG is a weekend; production RAG is a discipline. I have built the full retrieval stack from scratch twice, in two different enterprises, and the pattern that survives is layered: get ingestion and chunking right, make search hybrid, and put a knowledge graph where cosine similarity loses structure.
> what this looks like in practice
▸Engineered GraphRAG on Neo4j for the AI-Infused QE Platform with entity normalization at ingest, so "login page", "sign-in screen", and "auth UI" resolve to one entity instead of three retrieval misses.
▸Layered Agentic RAG over Jira and Azure DevOps so test generation is scoped against linked stories, epics, and acceptance criteria, not just similar-looking text.
▸Designed an enterprise conversational RAG pipeline end to end: chunking with overlap calibration over ~1,700 internal documents, hybrid search combining semantic and keyword signals (OpenSearch), and metadata filtering tuned for enterprise-jargon queries. Response relevance improved 30-40% in pilot; repeat escalations fell ~40%.
▸Guardrails sit at both ends: input-side scope checks and output-side grounding checks, so the pipeline fails safe instead of failing confident.
> where it shows up
The AI-Infused QE Platform (GraphRAG on Neo4j for relationship-heavy QA docs) and the enterprise conversational RAG migration (hybrid search over ~1,700 documents behind a load balancer). WealthOS deliberately inverts the pattern: event-sourced memory where retrieval would hallucinate.