Yuvraj Singh

From-scratch PyTorch replications of classic and SOTA AI/ML papers: transformers, attention, GANs, diffusion, CLIP, LoRA, MoE and more, with reproducible training code.

Educational distributed training and inference library for heterogeneous local hardware: Mac minis, Raspberry Pis and GPUs over plain Python sockets. FSDP, GRPO, parameter-server, expert and data parallelism.

A collection of my from-scratch implementation of various models deployed on HF Spaces

Trained a Whisper model a ~30M (whisper tiny.en) architecture I coded from ground up to build a small ASR model, going through the below-mentioned stage from scratch. Trained on GigaSpeech dataset form HuggingFace consisting of 250 hours of clean audio for a total of full 3 epochs

Distributed storage for ML checkpoints: shards .safetensors across a Raspberry Pi cluster over TCP with SHA-256 verification, replication and zero-config mDNS discovery.

TermDraw-Bench: a benchmark for whether LLMs and VLMs can generate and edit structured ASCII diagrams. 80 tasks across 4 categories, scored by a DeepEval LLM judge.