Luis Romero-Sevilla is vice president of artificial intelligence at Orbis Operations and the creator of Extended Cache Augmented Generation (ECAG), an architecture for answering questions across interconnected document collections that change too quickly for conventional retrieval systems.
He studied computer science and engineering at the University of Puerto Rico, Mayagüez, and pursued graduate training in artificial intelligence at Johns Hopkins University. At Digheontech, he built public-health software, including pandemic-data dashboards and patient-record infrastructure. Subsequent engineering work at Lockheed Martin and MIT Lincoln Laboratory encompassed data infrastructure and applied AI. He joined Orbis in 2024 and advanced into AI leadership, overseeing strategy, platform architecture, retrieval systems, and agentic workflows. His professional portfolio also describes an applied-AI research initiative linking Orbis with the University of Puerto Rico.
- Extended Cache Augmented Generation (ECAG): Romero-Sevilla distributes documents among multiple cached contexts operating in parallel. A supervising model questions each context, pursues promising connections, and synthesizes an answer spanning the collection. His AI Engineer presentation applies the design specifically to densely interconnected information that becomes obsolete quickly.
- Retrieval strategy follows data volatility: Vector-based retrieval updates quickly but can discard documents whose importance emerges only in relation to other records. GraphRAG captures those relationships, but repeated entity extraction and graph reconstruction become expensive when collections change continually. Romero-Sevilla's RAG and GraphRAG demonstration explores these competing approaches.
- Avoid rigid topical silos: Neatly categorizing cached contexts can cause a supervising model to skip apparently unrelated material containing crucial evidence. Romero-Sevilla instead favors balanced context buckets and iterative follow-up questions that expose unexpected relationships.
- KV-cache lifetime optimization: Cached context consumes computing resources, so retaining it requires explicit trade-offs among accuracy, latency, document turnover, and operating cost.
His other public code includes mystery-case.
