Edge-GNN: Constraint-aware graph neural networks for biological interaction modeling under edge deployment constraints (computational oncology, PPI networks).

This repository contains a set-associative cache simulator in C with four eviction policies: LRU, LFU, Bélády's OPT, and an online perceptron-based policy that scores by recency, frequency, and per-PC weights updated from a circular eviction history to explore whether a lightweight ML model can approximate OPT at runtime with no offline training.

Comparative analysis of Mamba vs. Transformers trained from scratch. Benchmarking Mamba's linear O(N) scaling and constant-time inference against quadratic attention mechanisms.

Neural network built from scratch in NumPy to classify data-center workloads based on thermal load indicators (CPU utilization, runtime), using Google cluster traces.

Repo for CUDA C++ GPU kernels for ML and HPC.

This repository is a CacheLib prototype of TARDIS, a learned cache eviction policy as part of the LDOS Expeditions project at UT Austin. It implements LRUForgive and S3FIFOForgive with a log-structured multithreaded architecture, and is evaluated against LRU & S3FIFO on real traces to to measure miss-ratio improvement under high eviction pressure.