Detail of Michelangelo's Creation of Adam: two outstretched hands almost touching

    Exploring AI from the systems side: how models, infrastructure, data, evaluation and reliability fit together.

    I learn infrastructure by running it myself, and breaking it as well. Most of what I've learned comes from my homelab and, thankfully, from the open source contributions I've made along the way.

    Pragalva Sapkota

    Introduction

    Hello everyone, welcome to my site! I'm Pragalva Sapkota, currently pursuing my undergraduate degree in Nepal. I find myself wondering how systems work and break, so most of my work revolves around that. I maintain a homelab, which is where I started learning the systems side of things, and I'm still exploring the field. Now I'm really interested in dissecting machine learning too.

    I find myself revolving around systems and ML, and I don't want to view either as a black box. Other than that, I'm a firm believer in community and love supporting open source projects whenever I can contribute meaningfully. For now, I'm just exploring tech and letting the tides take me on their journey.

    I · Experience

    Full CV →
    Sep 2026 – PresentInfrastructure Initiative LeadR&D Department, DWIT
    Mar 2026 – May 2026LFX Mentee, Schema and API GovernanceMeshery (CNCF)

    II · Projects

    All projects →
    devata · Talos Kubernetes homelabA Talos Kubernetes cluster on second-hand hardware, managed entirely from one public repo and reconciled by Argo CD. Every hour it publishes a snapshot of its own state; the live page renders from that file, stack and all.
    gpt-from-first-principlesAn executable learning path that builds a small GPT one concept at a time. The notebooks favor small tensors, visible intermediate values, and direct implementations before reusable PyTorch modules.
    embedding-models-from-scratchWord embedding models rebuilt from their papers in PyTorch, in order: Bengio 2003, word2vec, GloVe and ELMo. The ELMo notebook keeps its recorded A100 runs.
    jetson-nano-yolo-benchYOLOv8n on a Jetson Nano 4GB through TensorRT, with numpy pre/postprocessing, a ctypes CUDA runner, and every stage benchmarked: 38.6 FPS for the network, 12.3 for the live system.

    III · Recent writing

    All posts →