The Short Version

I’m an AI systems engineer progressing from data engineering through LLMOps and GPU engineering toward RL infrastructure — the most complex distributed systems problem in AI today.

This is not a story of pivots and detours. It’s a deliberate, logical progression toward a single, ambitious goal. Each layer builds on the last.

The Four Layers

My career maps to a 4-layer stack. Each layer solves a harder problem, and each one requires everything from the layers below it.

How Each Layer Builds

Think of it as building a skyscraper. Each stage adds a more specialized layer on top of a solid foundation.

Layer 1 → Layer 2: Data Engineering → LLMOps

My foundational skills in building robust, scalable data pipelines directly applied to the unique challenges of LLMs:

• ETL pipelines became the foundation for RAG and embedding pipelines
• Data quality and monitoring became the foundation for LLM observability
• Scalable systems knowledge enabled reliable inference APIs and agentic workflows

Layer 2 → Layer 3: LLMOps → GPU Engineering

This is the pivot from the application layer to the hardware layer — moving from using the tools to understanding how they work. It’s driven by asking “why?”:

• “Why is my vector search slow?” → Because the indexing algorithm causes non-coalesced memory access on the GPU, and the kernel isn’t optimized for this hardware’s cache size.
• “How can I serve more concurrent users?” → By implementing paged attention and speculative decoding at the kernel level.

LLMOps gives you the problems. GPU engineering gives you the solutions.

Layer 3 → Layer 4: GPU Engineering → RL Infrastructure

RL Infrastructure is the final and most important leap. It’s the most complex distributed systems problem in AI, and it requires mastery of all previous layers:

• RL requires massive data pipelines for experience replay, often petabytes in scale — a Data Engineering problem
• RL requires serving multiple models (actors, critics, world models) with different performance characteristics — an LLMOps problem
• RL requires extreme performance optimization to make training feasible — a GPU Engineering problem

My Journey: From Data to Decisions

My career has been a deliberate progression up the stack, from the fundamentals of data to the frontiers of AI infrastructure. Each step has been driven by a desire to solve harder problems and build more capable systems.

The Foundation: Data Engineering

My career began in data engineering, where I learned the fundamental principles of building robust, scalable, and reliable systems. I spent five years designing and implementing the data pipelines that power large-scale applications, mastering the art of moving and processing massive datasets. This experience taught me to think in terms of systems, not just code.

The Application: LLMOps

As large language models emerged, I pivoted to LLMOps, applying my systems-thinking to the unique challenges of deploying and managing AI in production. At Measurebit, I built the infrastructure for agentic workflows and LLM observability, learning how to bridge the gap between research and real-world application. This is where I first encountered the deep infrastructure challenges that limit the potential of AI.

The Deep Dive: GPU Engineering

To solve these challenges, I know I need to go deeper. I’m now studying GPU architecture, CUDA programming, and distributed training paradigms — moving from using the tools to understanding how the tools themselves work at a low level. This is the pivot from the application layer to the hardware layer, driven by a need to optimize performance at the most fundamental level.

The Frontier: RL Infrastructure

The end goal is the most complex distributed systems problem in AI: Reinforcement Learning Infrastructure. Training a million robots to learn a new skill requires mastery of the entire stack, from the high-level RL algorithm down to the low-level CUDA kernel. This is where every layer of my career converges, and I believe it’s where the most important work in AI will be done over the next decade.

Currently

• Senior Data Engineer @ Measurebit — building data and AI systems
• MTech in Data Engineering @ IIT Jodhpur — GenAI + RL specialization
• Writing about what I learn as I progress through each layer
• Currently at Layer 2 (LLMOps), building toward Layer 3 (GPU Engineering)

Get in Touch

I’m always happy to connect with others working on AI infrastructure, distributed systems, or the path from data engineering to RL.

• GitHub: github.com/saurabhghatnekar
• LinkedIn: linkedin.com/in/saurabhgghatnekar
• X/Twitter: x.com/saurabh_works