Search papers, labs, and topics across Lattice.
SecMate, a multi-agent virtual customer assistant (VCA), was developed for cybersecurity troubleshooting by integrating device-level diagnostics, user proficiency inference, and context-aware service recommendations. Device-level evidence from a local diagnostic utility significantly improved correct resolution rates from 50% to over 90% compared to an LLM-only baseline. User studies with 144 participants demonstrated high relevance of the recommender system (MRR@1=0.75) and a strong user preference for SecMate over human IT support at a lower cost.
Forget generic chatbots – SecMate slashes cybersecurity troubleshooting failures by 40% simply by adding device-specific diagnostics.
Recent advances in large language models and agentic frameworks have enabled virtual customer assistants (VCAs) for complex support. We present SecMate, a multi-agent VCA for cybersecurity troubleshooting that integrates device, user, and service specificity from conversational and device-level signals. Device specificity is provided by a lightweight local diagnostic utility, while user specificity relies on implicit proficiency inference and profile-aware troubleshooting. Service specificity is achieved through a proactive, context-aware recommender. We evaluate SecMate in a controlled study with 144 participants and 711 conversations. Device-level evidence increased correct resolutions from about 50% to over 90% relative to an LLM-only baseline, while step-by-step guidance improved pleasantness and reduced user burden. The recommender achieved high relevance (MRR@1=0.75), and participants showed strong willingness to substitute human IT support at costs well below human benchmarks. We release the full code base and a richly annotated dataset to support reproducible research on adaptive VCAs.
