Aananth V
Software engineer interested in systems, networking, and machine learning.
I’m currently pursuing a master’s in computer science at UT Austin. Before that, I spent four years at Google working on production RPCs, Cloud TPU networking, and the Linux networking stack.
Experience
Senior Software Engineer
I worked on the Prod RPCs team, developing the foundational Stubby and gRPC stacks that carry more than 95% of Google’s exabyte-scale network traffic.
Cloud TPU networking. Improved training and inference goodput by approximately 50–100% through changes spanning XLA collectives, a new gRPC transport, and the Linux networking stack.
TCP visibility. Tech led Fathom, a TCP observability effort for high-performance workloads; upstreamed Linux kernel telemetry and helped identify and fix a TCP Fast Open bug in mainline.
Open-source gRPC. Contributed vulnerability fixes, performance optimizations, and telemetry improvements to the gRPC repository.
Reliability and modernization. Led a two-year removal of a legacy streaming protocol and rebuilt RPC benchmarking, reducing test flakiness from roughly 10% to under 2%.
Samsung R&D Institute
Research and software engineering intern
Developed a knowledge-graph recommendation system using graph convolutional networks, and an earlier simulator for DAG-based blockchain protocols.
Arcesium
Software engineering intern
Built a configuration-driven search framework that reduced the time needed to deploy new search functionality by 20×.
Education
University of Texas at Austin
MS in Computer Science
National Institute of Technology, Tiruchirappalli
BTech (Honors) in Computer Science · Institute Gold Medalist · Rank 1/119
Publication
Transformer Based Motion In-Betweening
With Pavithra Sridhar, Madhav Aggarwal, and R. Leela Velusamy. A transformer-based system for synthesizing 3D animation between sparse keyframes.
ACCV Workshops, 2022Writing
How LLMs Work - Part 1: Attention
A from-first-principles derivation of attention, queries, keys, values, and the mechanism at the heart of modern language models.
RoPE and Local Attention hinder LLM performance
An experiment on why answer quality falls across long prompts — and what local attention and rotary embeddings may have to do with it....