Azka Baksh

I’m a research intern at LUMS working at the intersection of machine learning, reinforcement learning, and applied modeling. My focus is building models that are both technically strong and practically interpretable for real-world decisions.

In my education research, I developed agent-based simulations to study parental decision-making and school enrollment patterns, defining behavioral rules and supporting model validation and policy scenario analysis. In my engineering research, I designed RCMACA, a role-conditioned cooperative MARL framework with transformer critics and LoRA adapters to improve coordination on SMACv2.

I also work extensively across modern generative and representation learning—implementing PixelCNN, GANs with latent interpolation, and β-VAEs, and building diffusion models from scratch with forward/reverse processes and a denoising UNet. For multimodal learning, I implemented CLIP from scratch and built Transformer and Vision Transformer (ViT) models with custom patchification and self-attention.

Beyond core research, I build end-to-end applied systems: GNN-based agents for navigation, a spatio-temporal GCN for traffic prediction, an autoencoder-based MRI tumor segmentation model, and a Photosynth-style 3D reconstruction with an interactive virtual tour. I’ve also deployed a web-based inference service for face detection and gender classification.