Harsh Kumar Jha

Advisor: Prof. Shanmuganathan Raman, Head of CSE & EE, IIT Gandhinagar · Diffusion Models · Transformer-based Representation Learning · 3D Point Cloud Processing

• Designed and implemented an end-to-end latent-space diffusion framework for 3D point cloud denoising, centered on autoencoders to preserve global shape structure under synthetic noise.
• Proposed hierarchical transformer-based latent encoders with timestep conditioning, enabling diffusion over multi-scale geometric representations beyond raw coordinate-space diffusion.
• Achieved state-of-the-art denoising performance, improving Chamfer Distance by up to 2× over strong diffusion baselines (CD ×10⁻⁴: 1.23 vs 2.28 at noise level 1% on 10k-point clouds), validated via controlled ablations.

StyleGAN2 · GAN · PCA · CLIP · PyTorch

• Reproduced GANSpace on StyleGAN2-FFHQ, applying PCA over the W-space latents of a pretrained GAN to extract interpretable edit directions, ranked via CLIP contrastive scoring against target attributes.
• Showed that PCA-discovered directions remain semantically entangled (the top "beard" axis behaved as a gender axis), tracing this to attribute correlations in the training data.

Advisor: Prof. Anirban Dasgupta · Python · C++ · BERT · FAISS

• Benchmarked 7 retrieval strategies (BM25, Word2Vec, BERT bi-encoder, cross-encoder reranking) on SciFact (BEIR); found that hybrid BM25+BERT score fusion achieved higher precision (0.0923 vs 0.0917) and recall than cross-encoder reranking at 8× lower latency (31ms vs 240ms).
• Implemented a custom HNSW approximate nearest-neighbour index from scratch in C++ with pybind11 bindings, supporting inner-product metric for cosine similarity and binary graph serialisation, achieving 98.2% top-10 result overlap with FAISS exhaustive search.

B.Tech. in Artificial Intelligence · CPI: 8.51

Class XII (Physics, Chemistry, Maths) · 87.4%

Class X · 93.5%