Interpreting and advancing abstract reasoning models
We mechanistically interpret efficient and performant architectures on Abstraction Reasoning Corpus and build on top of them. ARC AGI is a benchmark of puzzles that evaluates an AI model's reasoning capabilities that are trivial for humans but hard for current data-hungry AI systems.
High-performance GPU kernels for AI workloads
triton::cu is a project done in collaboration with KLA. The aim of the project is to make an agentic pipeline capable of converting Triton code into CUDA code. CUDA code gives the programmer lower level control and therefore the code can be optimized more than Triton code.
Building an intelligent conversational system for low-resource languages by combining retrieval-augmented generation, translation, and speech processing. KathAI aims to make collected stories and language resources searchable and accessible while exploring audio cleaning, transcription, alignment, and speech-based interaction.
