Projects

• SpaLLM - Unified compressive adaptation of LLMs with sketching. Fine-tune compressed models directly using parameter-sharing sketches.
• SpartanServe - Fast concurrent LLM adapter serving with structurally sparse adapters, using custom Triton kernels and CUDA graphs.
• LLM Finetuning Project - Evaluated PEFT methods and 4-bit quantization for aligning small LLMs (Falcon, Gemma, Phi-2) on domain-specific tasks.
• NoSQL Document Database - A network accessible document database written in Golang with RESTful queries, updates, and subscriptions.