Hi there, I'm Dara 👩👋
I am a machine learning research engineer.
My work and interests evolves around foundation multimodal models, large-scale pre-training, and inference optimization. I am currently deep-diving into HPC for deep learning, writing custom GPU kernels in Triton that leverage tiling, shared memory, and parallel execution to overcome the memory wall. My focus is on accelerating DL training and inference through IO-aware kernel design.
I am concerned about AI safety and interpretability so I occasionally do some mechnaistic intrepretability probing and write about some of my findings here
- Infuse audio: A framework for aligning audio representations with the embedding space of LLMs (multimodality)
- Ablate compliance: Finding jailbreak directions within the activation subspace of a LLMs
- Flash Attention and Diffusion Kernels in Triton: Highly performant, highly optimized flash attention kernels, linear attention and diffusion models kernels