Generative models · Paper reproduction
Understanding diffusion models by implementing DDPM
A PyTorch implementation of the core denoising diffusion probabilistic model, motivated by understanding image generation from first principles.
Personal website · Research & engineering
Hello, I’m Hardik.
I study Mathematics and Computing at Manipal Institute of Technology. I’m interested in how intelligent systems learn, represent information, and behave—and in understanding why they work, where they fail, and when we can trust them.
I like moving between mathematical ideas and practical implementation: reading papers, building models, running experiments, and examining the results. My experience includes faculty-supervised computer-vision research, generative modeling, reinforcement learning, information retrieval, and quantitative software.
I’m increasingly interested in AI safety and mechanistic interpretability. I’m still exploring where those questions connect most naturally with my work.
Research
I’m a lead student researcher working with Dr. Ujjwal Verma on object detection and semantic segmentation in satellite imagery. The project is ongoing and subject to an NDA; I share only high-level methodology and approved information.
Selected projects
Stochastic processes · Quantitative engineering
LambdaLab: modeling trade arrivals with Hawkes processes
An exploratory quantitative-modeling application connecting point-process theory, parameter fitting, simulation, and interactive visualization.
Reinforcement learning
Learning through interaction: Q-Learning and PPO
Experiments with tabular value learning and neural policy optimization across several Gymnasium and Unity environments.
Information retrieval · RAG
Document retrieval and question answering
A modular document-search implementation built to understand how document representation and retrieval influence evidence-grounded question answering.
What I’m thinking about
How do useful representations emerge from training? What evidence tells us a model has learned something robust? How can we distinguish convincing outputs from trustworthy behavior? These questions connect my practical work to my growing interest in interpretability and AI safety.
Writing and notes → (forthcoming)