supraj.dev

Supraj Maripeddi · AI, Cloud & Platform Engineer

Cloud foundations. DevOps thinking. Useful AI.

Practical notes on cloud, DevOps and AI: the decisions, trade-offs and working code behind real systems.

Read the articlesWatch a short23:20
Articles
29
Handbook chapters
120
Interview questions
1,102
Years in production
10+

Latest from the notebook

01Latest writing

Fresh from the notebook.

Practical, long-form explainers on AI agents, GPUs, Kubernetes and the cloud.

All 29 articles

AI & Agents13 min read

What Is Google AX? How It Works with Agent Substrate

Understand Google AX through a shared office: tasks, workspaces, runners, and how Agent Substrate provides the sandbox. Follow one simple README check.

Agent Runtimes · Part 3

GPUs & Inference35 min read

GPU Architecture, Layer by Layer

A plain-language deep dive into a modern GPU, using the NVIDIA H100: GPCs, TPCs and SMs, every unit inside an SM, the memory system, occupancy, Hopper's clusters and TMA, copy engines, NVLink and NVSwitch, MIG, the graphics units, and where TFLOPS numbers come from.

Masterclass · Part 3

AI & Agents17 min read

What Is an Agent Harness? The Part of an AI Agent That Isn't the Model

Part 1 of Agent Harness Engineering. Why the same model can be a chatbot or a reliable coding agent, the seven parts every harness has, one bug fix traced through all of them, and the lessons Anthropic and OpenAI published about building harnesses.

Agent Harness Engineering · Part 1

AI23:20Narration: AI voice

NEW · SHORTS

One idea.
Watch it move.

Animated explainers that show the mechanism moving. Watch one, keep three learning points, then read the deep dive.

Now playing

LLM inference, end to end

  1. Your question makes ten stops. Only one of them, the inference engine, runs on a GPU; the rest run on ordinary CPUs. Three caches can skip work along the way.
  2. Prefill reads your whole prompt in one parallel pass and decides the time to first token; decode then writes one token per step, reading the KV cache each time.
  3. A 503 usually means no free capacity: no ready replica, full queues, or a full KV cache. It rarely means the model is broken.
Open the lessonAll shorts (4)

More shorts

03Series

Learn it in order.

Each series builds one idea at a time. Start at part 1.

04Library

Four free handbooks. 120 chapters.

Free handbooks you can read chapter by chapter.

Open the Library

05Practice

Practise until it clicks.

Interview questions with model answers, learning paths, knowledge maps and browser tools.

Hands-on courses

In preparation.

Articles, guides and roadmaps stay free to read.

06About

Hi, I’m Supraj. I build the platforms under useful AI.

I build the infrastructure that makes AI useful in production: RAG pipelines, agents connected to real systems, and the cloud platforms that keep them secure, observable and cost-effective.

My career spans more than a decade of cloud and operations engineering across AWS, Google Cloud and OCI. This is where I share the work, the decisions and what I learn along the way.

See my workConnect on LinkedIn