---
tags:
- machine-learning
- ml-systems
- robust-ml
- llm-inference
- computer-vision
---

<div align="center">

# Aman Singh

### ML Systems Engineer · Robust ML · GPU Inference · LLM Systems

I build reliable machine-learning systems across training, evaluation, inference, and secure deployment.

[GitHub](https://github.com/amanyagami) · [Portfolio](https://aman-ai-portfolio.netlify.app) · [Résumé](https://asinghresume.netlify.app) · [LinkedIn](https://www.linkedin.com/in/amanyagami)

</div>

## Hub highlights

| Type | Resource | What it shows |
| --- | --- | --- |
| **Space** | [Viyog OOD-vs-ADV Leaderboard](https://huggingface.co/spaces/amanyagami/viyog) | Interactive evaluation for separating adversarial and out-of-distribution inputs |
| **Dataset** | [viyog-adversarial](https://huggingface.co/datasets/amanyagami/viyog-adversarial) | Adversarial/OOD evaluation data for reproducible robustness experiments |
| **Model** | [Cifar100_Finetuned](https://huggingface.co/amanyagami/Cifar100_Finetuned) | Fine-tuned vision model checkpoint and evaluation artifacts |
| **Model** | [viyog-weights](https://huggingface.co/amanyagami/viyog-weights) | Viyog research weights for robust input-separation experiments |
| **Model** | [ViT-Base checkpoint](https://huggingface.co/amanyagami/Vit_base_patch16_224.pth) | Vision Transformer checkpoint for downstream experiments |

## Research

### Viyog: Separating Adversarial and Out-of-Distribution

**Accepted at ESWEEK CODES 2026.**

Viyog studies how intermediate-representation geometry can separate adversarial inputs from out-of-distribution inputs. This Hub profile contains the accompanying models, dataset, and live evaluation Space.

## Engineering focus

- **Robust ML:** adversarial evaluation, out-of-distribution detection, representation geometry, and reproducible PyTorch experiments.
- **GPU systems:** H100 training with DDP and FP16/BF16, FSDP checkpointing, and sub-10 ms A6000 inference.
- **LLM infrastructure:** model integration, streaming tool calls, bounded concurrency, vLLM rollout networking, and evaluation.
- **Secure execution:** Rust sandboxing, Windows AppContainer/LPAC confinement, and Linux filesystem isolation.

## Open-source systems work

Contributions across [SGLang](https://github.com/sgl-project/sglang), [VERL](https://github.com/verl-project/verl), [Megatron-LM](https://github.com/NVIDIA/Megatron-LM), [PyTorch AO](https://github.com/pytorch/ao), [LLVM/MLIR](https://github.com/llvm/llvm-project), and [arapuca](https://github.com/sergio-correia/arapuca), including KV/radix-cache correctness, FSDP checkpoint integrity, pipeline-parallel execution, FP8/quantization, compiler correctness, and sandbox isolation.

## Background

- **Arizona State University, MPS Lab** — Graduate Researcher · Tempe, Arizona
- **Amazon** — Software Development Engineer · Bangalore, India

## Education

- **M.S. in Computer Engineering** — Arizona State University, 2024–2026
- **B.Tech in Electrical and Electronics Engineering** — National Institute of Technology Karnataka, 2018–2022

## Toolkit

**PyTorch · CUDA · DDP/FSDP · SGLang · vLLM · Transformers · Python · C++ · Rust · Go · Linux · Docker · Kubernetes · AWS · Google Cloud**

## Connect

For robust ML, GPU systems, LLM infrastructure, or research engineering: [email me](mailto:2amansingh2@gmail.com) or connect on [LinkedIn](https://www.linkedin.com/in/amanyagami).