Shantanu Kumar

Vision-Language Transformer

A PaliGemma-style Vision-Language Transformer implemented from first principles, including ViT, contrastive learning, RoPE, GQA, KV-cache, and autoregressive decoding.

Foundation LLM Pipeline

Building an end-to-end LLM pipeline from a base Transformer through supervised fine-tuning, reward modeling, and PPO-based RLHF.

Sythra.ai (Previously TheKourse)

AI-powered interactive learning platform with contextual agent explanations, automated project scoring, and in-browser code execution engine. Secured funding from Italy.

Vanilla PyTorch Transformer

Implemented the original Transformer architecture from scratch in PyTorch to gain a deep understanding of self-attention, masking, and encoder-decoder stacks.

Agentic Blog Generation

An agentic blog generation system using LangGraph and FastAPI, orchestrating multi-step LLM workflows with tool use, memory, and validation pipelines.

ResNet-34 Audio CNN

A ResNet-style CNN for audio spectrogram classification with end-to-end inference deployment using PyTorch, torchaudio, FastAPI, and Modal.

Multi-Agent Justice (VIDHI-AI)

Vidhi AI is a multi-agent courtroom simulation modeling legal reasoning through role-specific AI agents, grounded in statutory knowledge using RAG.

Gigabyte Phantasia

Build the landing page website for the 24 hour continous hackathon Gigabyte phantasia the design is inspired by the vegas theme.