Mechanistic Interpretability Implicit Personalization Monitoring and attributing the user models LLMs silently build. SPAR research fellowship MoE Interpretability Adapting HeadPursuit / SOMP to classify expert specialization in Mixture-of-Experts LLMs ML BayesianFlow Pixel-wise uncertainty estimation in Flow Matching generative models via Last Layer Laplace Approximation LoRA & DoRA in TinyGrad From-scratch Low-Rank Adaptation and Weight-Decomposed LoRA implemented in TinyGrad Vector Store + RAG Minimal RAG pipeline with a custom vector store and Mistral via Ollama. No LangChain HPC Self-Attention Kernels Optimized Causal Multi-Head Self-Attention in CUDA, OpenMP, and SIMD. 1.09× faster than PyTorch naive on A100 Parallel Heat Stencil Efficient and scalable 5-point heat stencil in C, parallelized from one core to multiple nodes with hybrid MPI + OpenMP on Cineca Leonardo
