Kernels
&
Tensors

Exploring the depths of software engineering, one layer of abstraction at a time.

The Profile

I am a Computer Science undergraduate driven by a relentless curiosity for how systems work under the hood. I believe in understanding the fundamentals - from the math powering machine learning algorithms to the memory management of low-level software. My focus is on writing clean, robust code and building architectures that scale.

// kush_singh

Selected Work

Systems Engineering

Inference Engine

High-performance C++ inference engine for LLMs using SIMD/AVX2, with KV-Cache and quantization strategies. Built for raw speed on edge devices - optimized at the silicon level.

C++ SIMD/AVX2 KV-Cache Quantization

Source

LLM Fine-tuning

Mental Health ChatBot

Fine-tuned Llama 3 8B on counseling datasets to enhance empathetic reasoning using PEFT adapters. Compressed via 4-bit Quantization (GGUF) for efficient edge deployment.

Llama 3 8B PEFT/LoRA GGUF PyTorch

Source HF ↗

Deep Learning

Transformer Translate

Built from scratch with mixed precision. Achieved BLEU 23.64 (En→Hi).

PyTorch Transformers AMP

Source Live Demo ↗

Web Systems

Kosh SSG

Custom static site generator in Golang with Markdown-to-HTML conversion & KaTeX support.

Golang Markdown KaTeX

Source Live Demo ↗

Computer Vision

Image Captioning

CNN + LSTM model for automatic image description generation, deployed on HuggingFace Spaces.

PyTorch CNN LSTM

Source Live Demo ↗

Education

Micrograd

A scalar-valued autograd engine built from scratch in Python - understanding the roots of backpropagation and neural networks at the fundamental level.

Python Autograd Backprop

Source