AI Security Researcher · Adversarial ML · Inference Optimization

Kathmandu, Nepal

Website · Google Scholar · LinkedIn · ResearchGate · Medium

I work on AI security and efficient inference, with a focus on adversarial behavior in language models and systems-level optimization.

• LLM safety and adversarial evaluation
• Multilingual and code-switched model behavior
• Activation-level analysis
• LLM inference optimization
• GPU kernel optimization
• Hardware and microarchitectural security
• ML compiler behavior

Nepali Adversarial Safety Benchmark

Research on safety behavior in LLMs under Nepali and code-switched inputs.

Paper

Training-Free Activation Sparsity for LLM Inference

Experimental inference-optimization framework using activation sparsity and custom CUDA kernels.

GitHub · PyPI

• Compiler Optimizations & Model Safety — investigating whether compilation and inference transformations affect safety-relevant model behavior.
• Transformer Side-Channel Security — exploring potential microarchitectural side channels in transformer inference.
• Activation-Level Safety Analysis — studying activation patterns associated with safety-relevant model behavior.

Python · C/C++ · CUDA · PyTorch · Triton · MLIR · Linux