RAG & Enterprise Search

My retrieval skill is organized around a conviction the systems keep confirming: the answer quality of a RAG product is decided in the retrieval layer, long before a model writes a word. So that layer gets the engineering: chunking tuned to the corpus rather than a default window, embeddings chosen and swapped behind an abstraction, vector and full-text search run side by side and fused with reciprocal rank fusion, a cross-encoder reranker sharpening the fused top, and two-layer storage that matches on small semantic units while returning the parent document a reader actually needs. Queries get the same respect as documents — language detection, translation into the corpus language, complexity analysis, decomposition of compound questions, and multi-rewrite expansion, each as an explicit pipeline stage that can be observed and skipped rather than folded into one opaque prompt.

The proving ground is a shared enterprise retrieval platform that serves a company's knowledge corpus to support agents, analysts, developers, and other AI products through REST and MCP interfaces backed by one retrieval core — built as a platform deliberately, so every consumer inherits each retrieval improvement at once. The same discipline runs through my open-source work: a durable agent-memory server whose full-text, semantic, hybrid, and reranked retrieval modes are first-class and swappable, and — next — a benchmark that measures which of those modes actually earns its latency for agent memory workloads. Retrieval quality is watched rather than asserted: the pipeline's stages log their own score distributions, deduplication counts, and percentile bands, so quality is calibrated from production evidence; the reranking stage earned its place by raising relevance and cutting noise, and the benchmark I am designing next turns that calibration into measurement against ground truth. The systems are listed in the Related Projects and Related Experience panels above.