Manjit Pokhrel

I’m an independent AI security researcher building reproducible attack pipelines to stress-test models and evaluate alignment robustness in open-weight LLMs. I run experimental methods, document results, and disclose responsibly to vendors.

My work centers on adversarial ML and AI safety—especially prompt injection and jailbreak-style attacks. I’ve performed red teaming across Gemma, Qwen, and Ollama, identifying safety bypass behaviors like PII leakage and restricted code generation, and I submitted a responsible disclosure report to Google AI VRP.

I also develop practical deep learning and deployment projects, including a “MiniGPT” GPT-style transformer implemented from scratch (manual backprop, multi-head self-attention, positional encoding) and LLM benchmarking and fine-tuning workflows using HuggingFace and Unsloth with LoRA + CUDA acceleration. I enjoy turning research insights into working systems, including an AI-powered Discord recommendation bot using the TMDB API.