Projects

Custom built LLM - ongoing project. Code for pretraining, mid-training and post training an LLM (SFT & RL) on a HPC cluster. Implements experimenting, logging and evals.

Implements key PyTorch functionalities from scratch with the philosophy "What I cannot build, I cannot understand". Includes autograd, neural network layers, transformers, and ASR models.

BSc thesis exploring how transformers learn syntax. Training GPT-2 on formal grammars and probing hidden representations to uncover where hierarchical structure emerges.

Sample-efficient LLM fine-tuning for math reasoning: training on hard examples outperforms random selection under fixed compute budgets.

AI-powered news-to-podcast converter. Curate articles and generate personalized audio briefings in ~30 seconds with morning news mode and debate mode using Grok Voice API.

Examines social biases in State Space Models (Mamba) versus transformers across multiple demographics (race, gender, profession, religion), applying debiasing techniques with LoRA.