Bhooyas Kapadia

Developed a synthetic algebra dataset generator to create verifiable linear equation solving tasks for controlled reasoning evaluation. Established SFT baseline and applied Generalized Reward Policy Optimization (GRPO) with continuous numeric reward shaping and KL-regularized policy updates. Achieved 98% accuracy versus 19% (SFT) and 0% (base model), demonstrating significant reasoning gains through reward-driven post-training under 4GB VRAM constraints. Conducted stability analysis via reward tracking and entropy monitoring to prevent policy collapse.

Huggingface Link

Fine-tuned TinyLlama (1.1B) for improved instruction-following under strict hardware constraints (single NVIDIA GTX 1650 Ti, 4GB VRAM). Implemented parameter-efficient training using LoRA adapters, mixed precision (FP16), and gradient accumulation to remain within memory limits while maintaining stable convergence. Optimized LoRA rank and learning rate through controlled experiments on the Dolly 15k dataset, reducing trainable parameters by >99% compared to full fine-tuning. Achieved improved instruction coherence and structured response consistency.

Huggingface Link

Designed and implemented a custom transformer architecture from scratch using PyTorch, resulting in StoryNet—a specialized neural network for creative text generation. Successfully trained the model to generate coherent short stories, demonstrating expertise in transformer architecture, sequence modeling, and natural language generation.

Implemented a Conditional Variational Autoencoder (ConditionalVAE) that generates handwritten digits based on specified class inputs, trained on the MNIST dataset. Demonstrated the ability to control the generation process through conditional parameters, highlighting expertise in deep generative modeling, latent space manipulation, and complex neural architecture design for controlled image synthesis.

Engineered AvatarGAN, an advanced image generation pipeline combining Deep Convolutional GAN (DCGAN) with Super-Resolution GAN (SRGAN) to create high-quality game character avatars. The system generates characters at multiple resolutions, with DCGAN handling initial character creation and SRGAN enhancing image quality through upscaling.