LLMaven: Large Language Model based Agentic Verification and Exploration of Natural Inquiry

Research teams across domains are eager to integrate large language models (LLMs) into their workflows—but off-the-shelf tools often fall short when it comes to research-specific use cases, reproducibility and accessibility. At the Scientific Software Engineering Center (SSEC), we are building LLMaven, an open source platform to bridge that gap: an extensible multi-agent AI framework designed for scientific researchers who need more than a chatbot. LLMaven brings together containerized cloud-native tools, standardized agent protocols, and plug-and-play backends to form a secure, reproducible, and flexible system. At its core, LLMaven utilizes the Agent-to-Agent Protocol (A2A) and Model Context Protocol (MCP) to support structured communication between agents and external resources. It allows agents to reason collaboratively and dynamically delegate tasks across domains—whether querying temporal knowledge graphs, managing GitHub repositories, or coordinating custom pipelines. Built on Kubernetes with Helm for scalable deployment, the system incorporates tools like OpenWebUI for secure interactions, Grafana and Logfire for observability, and a hybrid data backend (Neo4j, Postgres, MinIO). Our architecture supports switching between local and remote models utilizing containerized inference engines and API-based backends without changing the application logic—making it future-proof and vendor-flexible. With support via an NSF and OSTP’s National AI Research Resource award, LLMaven has peer reviewed validation and access to compute resources. In this talk, we’ll walk through the architectural decisions, lessons learned from our user-centered MVP development, and a real-world use case featuring data from the Rubin Observatory project. We’ll also discuss our vision for democratizing AI tooling in research and how the LLMaven model could enable cross-domain collaboration at scale. LLMaven isn’t just a platform—it’s an evolving conversation between code, practice, and people. And that makes it a story worth sharing.