Vignesh Radhakrishna
Machine Learning Engineer, Meta Platforms · Menlo Park, California
About
I build applied machine learning systems for large language models, with a focus on the training, evaluation, and deployment infrastructure that lets frontier AI reach people at scale. My work is centered on agents and the path toward more general AI systems. I contributed to the development of Llama 3 at Meta, working across pre-training and post-training to improve model quality and the tooling used to ship it.
My work sits where research and production meet: turning ideas from the open research community into reliable systems that serve millions of users. I care about open, reusable AI infrastructure and its role in advancing US leadership in artificial intelligence.
Experience
Machine Learning Engineerat Meta Platforms
Contributing to the Llama family of large language models across pre-training and post-training. Building evaluation, training, and release infrastructure used by the Llama team.
Machine Learning Engineerat Meta Platforms
Worked on applied ML systems for ranking and retrieval, scaling model training pipelines and improving inference latency for products serving hundreds of millions of users.
Software Engineerat Earlier roles
Built distributed systems and data infrastructure supporting machine learning workloads at scale.
Education
M.S., Computer Scienceat University of Massachusetts Amherst
Coursework and research in machine learning and systems. UMass Amherst is a long-established hub for AI and information retrieval research.
B.E., Computer Scienceat RV College of Engineering
Foundations in algorithms, systems, and machine learning.
Research & Publications
Work spanning large language model training, evaluation, and the systems that support applied ML, with publications appearing in peer-reviewed venues including NeurIPS and IJECE. A full publication record and citation metrics are available on Google Scholar.
Research on machine learning methods, presented at NeurIPS
Applied machine learning work published in IJECE
Endeavour
I build frontier models, and the limiting factor in that work is data. The quality of what a model learns is bounded by the quality of what it is taught, so a meaningful part of my effort goes into curating, filtering, and refining the data that goes into training.
That means building pipelines that source large volumes of text and code, then remove the noise: deduplication, quality classification, safety filtering, and careful mixing across domains and languages. The goal is a corpus dense in useful signal, where every example moves the model toward better reasoning rather than memorizing repetition. Better data yields better models, and that is where the work is.
National Impact
Large language models are a cornerstone of US technological leadership, with broad national importance across industry, healthcare, defense, and scientific research. My contributions advance this national interest in several specific ways:
Direct technical contributions to Llama 3, an open large language model released by Meta and used widely across US industry, startups, and academic research to build and study AI applications.
Development of training and evaluation infrastructure that improves the reliability and reproducibility of frontier AI systems, supporting the broader US AI research and engineering community.
Work on applied ML systems serving hundreds of millions of users, demonstrating the ability to move research from prototype to production at national scale.
Advocacy and contribution to open, reusable AI tooling, lowering the barrier for US-based researchers and small companies to participate in frontier AI development.
A background in both machine learning and systems, from UMass Amherst and RV College of Engineering, positioning ongoing work to strengthen US leadership in the infrastructure underlying modern AI.