Divyanshu Pabia

Why computing matters to me

I started programming in Java in grade 7, and I loved the idea that I could describe a task precisely enough for a computer to carry it out. Computers became more than devices to use: they were tools I could understand, change, and build with. The internet opened up an enormous world of knowledge, while programming gave me a way to turn that knowledge into something useful.

My ICSE schooling gave me a strong foundation in mathematics and computer science before I completed grade 10 in 2019. Mathematics appealed to me as the language of science, and physics made me curious about the underlying rules of the world. Computing connected those interests through the ability to simulate, explore, and solve problems beyond what a person could calculate alone.

Wanting to look under the hood

I wanted more control over my development environment than clicking through menus could offer. In grade 8, I tried dual-booting a Hackintosh on a Dell Inspiron with a second-generation i3 and 2 GB of RAM. It ran terribly, but experimenting with it taught me that the system itself was something I could investigate.

I explored websites, Python programs, and JavaFX applications, then moved into Linux virtual machines, beginner CTFs, Hack The Box challenges, and cryptographic puzzles after grade 10. Linux gave me the control and visibility I had been looking for. That instinct still shapes my work: when something behaves unexpectedly, I want to understand what is happening beneath the abstraction.

Learning by building with people

During my B.Tech at SRM, Next Tech Lab's McCarthy Lab introduced me to supervised learning, NLP, and reinforcement learning for game-playing. Think Digital and Tech Analogy exposed me to application development and collaboration with people working in different technical domains. Hackathons pushed me to turn an idea into a working implementation under real constraints.

The Apple and Infosys iOS App Development Program led me into Swift and the Apple ecosystem. Building Gamergrid at the iOS App Development Centre, working as an app development intern at Infosys, and learning Java backend development gave me experience moving from an idea to a usable application. Docker, AWS, Terraform, and CI/CD helped me understand how that software reaches production.

Open-source work added another perspective. Contributing to projects such as OpenELIS and Tor meant reading an existing codebase and fitting my work into a larger system. As a remote contractor with Tor's Network Health Team, I contributed to reimplementing the Java Tor Metrics Library in Rust, working on retrieving, parsing, and validating network descriptors from CollecTor. I also worked on the descriptor pipeline connecting object storage and a dedicated parser service with PostgreSQL and VictoriaMetrics exports. The work made me think carefully about memory safety, extensible interfaces, and how reliable data flows support historical analysis and future real-time dashboards. I value that kind of engineering: understanding what already works, why it works, and how to improve it without losing its purpose.

Where I am heading

My current direction is AI Systems Performance Engineering, with a focus on large language models. I want to understand both how an LLM is built and what makes its training and inference efficient. That means connecting transformer architecture and numerical computation with memory hierarchies, GPU kernels, compilers, scheduling, and distributed execution.

I am developing this depth through C/C++, Python, PyTorch, CUDA C++, Triton, and JAX. My objective is to be able to follow a workload through the stack, identify its bottleneck, and justify an improvement with measurements. Over time, I want to contribute to inference and serving projects such as vLLM and SGLang, especially around memory use, kernels, and runtime performance.

My existing projects reflect the path that brought me here: backend systems, mobile applications, architecture experiments, and research. They are not all LLM systems projects, and I am building that part of my portfolio deliberately. I want future work to demonstrate understanding through implementations, reproducible benchmarks, and clear technical writing.

Graduate study and the longer-term goal

I completed my B.Tech in 2025 with a CGPA of 8.5/10 and am preparing for a master's intake in 2027, with NUS and NTU among my primary targets. I want graduate study to deepen my mathematical and systems foundations and give me the opportunity to work on research problems where algorithms, software, and hardware meet.

My long-term goal is to become an AI systems specialist who can help make large-scale computation more efficient and accessible. I am particularly drawn to the infrastructure behind modern AI: the runtimes, compilers, memory systems, and distributed machinery that turn a model into a system people can actually use. Research, open-source contributions, and a public record of careful engineering are the steps I am taking toward that goal.

How I learn and work

I learn best from first principles: work through the mathematics, implement a small version, inspect the source, debug it, and measure what happens. I enjoy understanding the details, but I have also learned that chasing completeness can delay useful progress. I am working on balancing depth with iteration: finish a version, test its assumptions, and improve it with evidence.

My local hardware keeps that approach practical. I use an M1 Max MacBook Pro for development and research and a GTX 1050 machine for CUDA experiments, with cloud GPUs when a larger experiment needs them. Limited resources encourage me to ask what an experiment really needs and what I can learn from a smaller, well-designed run.

Beyond the code

I was my school's sports captain, competed in swimming at national and inter-state events, and played football and athletics. From June 2020 to November 2022, I played competitive Call of Duty Mobile in Mumbai. Across my teams, we qualified for the World Championship in 2020 and 2021 and were supported by Force 1 Esports, Mindfreak, and Revenant Esports along the way.

I am competitive, and I care about doing things well. Sport and esports taught me the value of deliberate practice, teamwork, and staying composed when the result matters. I bring that drive to engineering, while learning to make the process sustainable and leave room for curiosity.

I would like to connect with people working on LLM systems, GPU computing, and performance engineering, whether through a research opportunity, an open-source contribution, or a conversation about a difficult systems problem.