Agentic AI · LLM security
Gourob Debnath
AI Software Engineer and Graduate Researcher
I design production multi-agent systems and study how they fail: memory, tool-use, and LLM security under real constraints.
Selected evidence
Published
IEEE ICCIT 2025
Empirical evaluation of comment-based contextual deception in LLMs. Attack success 63–93.5% across three models; supporting dataset on Zenodo.
Production
Sara, 500+ users
Multi-agent healthcare assistant with MCP tools, long-term memory, voice, and production tracing.
On-device
AssesGen
Local multi-agent IELTS generation and evaluation on Gemma and TTS, with no cloud APIs.
Current role
Software Engineer, SJ Innovation LLC
2024 – Present
Education
B.Sc. in Computer Science & Engineering
Metropolitan University, Sylhet · CGPA 3.54 / 4.00
Latest publication: Code Poisoning Through Misleading Comments: Jailbreaking Large Language Models via Contextual Deception (IEEE ICCIT 2025).