Harshavardhana Naganagoudar

▾ [About]

I'm an AI/ML engineer exploring reasoning models, autonomous agents, AI alignment, and interpretability. I enjoy building practical AI systems across reinforcement learning, LLM workflows, and multimodal applications, while studying how these systems reason and behave. After six years in software testing and automation, I transitioned into AI/ML to work on systems that combine intelligence, creativity, and real-world usability.

▾ [Projects]

Agent Alignment Evaluation

Can an autonomous AI agent remain honest, safe, and corrigible when deception and rule-breaking increase its reward?. This project evaluates autonomous AI agents in simulated business environments designed to create conflicts between an agent's objective and desirable behavioral constraints. GitHub

Reasoning Trace Corruption

Will the LLM model accept the provided plausible fake Chain of Thought with wrong answers or will it self correct and arrive at the right answers? GitHub

Opaque Reasoning Experiment

What if an AI could decide which parts of its reasoning you get to see—and hide the rest in plain sight? GitHub

▾ [Previously]

I previously worked at IBM as a Test Specialist, focusing on automation, backend testing, CI/CD pipelines, and release reliability for enterprise systems.

▾ [Interests]

I enjoy running, reading, anime, and puzzle games — anything that rewards curiosity and patience. A lot of my learning comes from exploring ideas online, especially listening to Andrej Karpathy somehow make neural networks feel surprisingly intuitive.