Conversational AI system over unstructured spiritual text. Converted 10K+ lines into an optimised SQL database, preserving semantic integrity. Deployed as a browser-based agent with on-edge processing for data privacy.
Fine-tuned a seq2seq translation model on 8000 Gujarati couplets of Bhaktchintamani to generate poetry in the style of Poet Nishkulanand Swami from English prompts. Deployed as a public web application.
Implemented GANs, VAEs, and Diffusion Models entirely in raw PyTorch, including DCGAN, conditional GAN, beta-VAE, VAE-GAN, DDPM, i-JEPA, Score Matching, and Implicit Sampling directly from research papers.
Led a team of 3 in a Kaggle-based competition to predict user geographical coordinates from tweet data. Deployed a fine-tuned model mapping user sentiments to specific geographical coordinates via NLP.
Proposed a novel function regression approach using neural networks that avoids backpropagation, exploring new learning mechanisms via unidirectional (forward-forward) training. Implemented Kolmogorov-Arnold Networks (KAN) to reduce parameter count while achieving MSE of 0.001.
Delivering client-facing demos and whitepapers in applied AI. Working across network intelligence, knowledge management, and autonomous agent systems.
Developed AI features for a food robotics product — from recipe scaling to agentic cooking command generation from custom inputs.
Applied machine learning to computational fluid dynamics, predicting intermediate timesteps and reducing simulation overhead for laminar flows.