writing about distributed systems, cloud architecture, and software engineering
notes on things i'm working on and learning about. mostly distributed systems and cloud stuff.
A lightweight, cloud-native Kubernetes Java client that eliminates configuration complexity by providing native support for AWS IAM authentication (EKS), GCP authentication (GKE), and Azure authentication (AKS) - without requiring kubectl or complex kubeconfig management.
modern tui for kubernetes with vim-style keybindings, tuned for ai/ml clusters
eBPF-powered observability toolkit for Kubernetes AI/ML workloads with deep visibility into container networking, syscalls, and GPU performance
Containers are just Linux processes with isolated network stacks. Understanding how packets actually move between them is the difference between building observability that works and building observability that lies to you.
You wrote 200 lines of Rust. The binary is 47MB. The eBPF verifier rejects your program for being too large. Without understanding linking and memory layout, you're guessing. LLD maps show you exactly where every byte came from.
Traditional monitoring kills performance. eBPF lets you observe everything at the kernel level without slowing down—or breaking—anything.
I'm building what might be the world's largest internal deployment system. In 2025, we expanded from bare-metal to cloud-native resources—dynamic fleets, tag-based targeting, and autoscaling integration.
Building infrastructure that serves thousands of customers from shared resources is deceptively hard. Here are the actual problems and battle-tested solutions.
AI tools made building frontends trivial, exposing that backends were never simple. Here's why your AI-generated app can't deploy, and how platforms like Lovable Cloud actually solve it.
I joined a tiger team to build the company's first large-scale internal language model. My job? Build the data pipelines that could feed an 8 billion parameter model. Here's what happens when data engineering meets cutting-edge AI.
When automation systems need to know if they can touch a resource, 'is this production?' becomes a surprisingly complex question. I built a service to classify infrastructure by production confidence scores.
How do you deploy security patches across thousands of compute instances spanning multiple regions without causing an outage? Infrastructure-as-code breaks at scale. I built a global patching system to handle it.
What happens when you give your deployment system the power to say 'no'? I built an auto-rollback service that detects anomalies in real-time and reverses bad deployments. Here's why accuracy is harder than you think.