Building Safer AI: Alignment and Robust Cyber-Physical Systems
Why the next generation of AI must be predictable, aligned, and physically grounded
Dec 4, 2025
Kundan Kumar
AI Safety Research Fellow @ Algoverse | PhD Researcher | Safe Reinforcement Learning, Multi-Agent Systems & LLM Agents
Hi! I’m Kundan Kumar, a Ph.D. candidate in Computer Science with a minor in Statistics at Iowa State University and currently an AI Safety Research Fellow at Algoverse. My research centers on building safe, reliable, and adaptable AI systems for next-generation cyber-physical infrastructure, including smart grids, autonomous systems, and multi-agent environments, with a particular focus on evaluations, adversarial robustness, and scalable oversight for agentic systems.
I design safety-critical deep reinforcement learning (DRL) systems that integrate domain knowledge, uncertainty, and constraints for robust decision-making under distribution shifts and partial observability. My focus includes adversarial robustness, transfer learning, and scalable oversight for reliability in high-stakes environments. Recently, I’ve developed LLM-integrated frameworks that connect perception, planning, and language reasoning, linking low-level control with interpretable decision-making. I am particularly interested in AI safety, alignment, and evaluation at the intersection of foundation models and physical systems.
Beyond research, I enjoy sharing my insights through educational content on Substack and YouTube. Outside of work, I love cooking and Ice skating 🛼.
Other Research Interests
AI Alignment & Safety
Agentic evaluations, adversarial robustness, scalable oversight, and alignment for foundation model–based systems. LLM Reasoning & Agents
Agentic workflows (LangChain/LangGraph), tool-use, memory, retrieval, planning & reflection loops for robust multi-step reasoning. Autonomous Perception & Control
Vision-based perception (detection, segmentation, sensor fusion) integrated with learning-based control and trajectory planning for autonomous systems.
Agentic evaluations, adversarial robustness, scalable oversight, and alignment for foundation model–based systems. LLM Reasoning & Agents
Agentic workflows (LangChain/LangGraph), tool-use, memory, retrieval, planning & reflection loops for robust multi-step reasoning. Autonomous Perception & Control
Vision-based perception (detection, segmentation, sensor fusion) integrated with learning-based control and trajectory planning for autonomous systems.
Explore My Work
Talks
Publications
News Highlights
[Jan 2026]
Selected as an AI Safety Research Fellow at
Algoverse.
Developing and evaluating methods for agentic AI safety, including robustness testing, oversight, and evaluation frameworks.
[Dec 2025]
Completed an AI Strategy, AI Safety, Biosecurity program hosted by
BlueDot Impact
, focusing on long-term AI risk, governance, and responsible deployment of advanced AI systems.
[Sep 2025]
