Hello! I am Ajinkya Kiran Mulay, a Research Scientist at Meta. I defended my Ph.D. in 2024 from the Electrical and Computer Engineering Department at Purdue University under the guidance of Prof. Xiaojun Lin. My Ph.D. research focuses on developing theoretical results for exact basis recovery in the context of Differential Privacy and Federated Learning for sparse models in the under-determined domain. My other research work spans the fields of Computational Social Sciences, Adversarial Robustness, and their relation to privacy. Some of my notable research/industry experiences have been at Facebook (Meta) and the University of Tokyo.
Please find the list of my publications under the research page and my CV under the CV page. If you’re curious about my other software projects, check out the curious page.
Updates
Will be attending the Workshop on Assurance and Verification of AI Development (AVID) on May 17th, co-located with IEEE S&P in San Francisco, organized by FAR.AI in partnership with the Center for AI Safety.
Attended ControlConf from April 17–19, 2026.
Received third prize in the Sentient Arena Cohort 0 OfficeQA Leaderboard challenge, building state-of-the-art skills for coding agents to beat OfficeQA, an enterprise-relevant benchmark by Databricks over SEC filing data.
Placed 6th at AgentCTF × AgentXploit, organized by the Berkeley RDI Center — a security research competition where participants build AI agents that autonomously identify and exploit real-world CVEs across frameworks like LangChain and AutoGPT. Achieved a full score of 3.55 across the 20-task evaluation set, running exclusively on Claude Haiku.
Promoted to Senior Research Scientist at Meta.
We won the XAI Hackathon (demo) under the Grokipedia track for real-time updates to Grokipedia from latest X.com posts.
My personal interests rely on developing small efficient frontier models that can capture human reasoning capabilities with parameters that can easily run on most home hardware. Most of my topics will focus on such architectures.
Started studying AI safety deeply, mainly the Technical AI safety course at BlueDot Impact. Happy to chat more.
I officially defended my PhD on the topic of "Differentially Private Federated Learning Algorithms for Sparse Basis Recovery."
Our manuscript "SPriFed-OMP: A Differentially Private Federated Learning Algorithm for Sparse Basis Recovery" is accepted for publication in the Transactions of Machine Learning Research (TMLR) journal.
We recently added a preprint on sparse basis recovery guarantees in the Differential Privacy-Federated Learning (DP-FL) domain. We demonstrate that even with a few samples (n), where the data dimensions (p) are high (i.e. p >> n), we can recover the exact sparse model with high probability. This is the first theoretical work that proves such guarantees with limited samples under the DP-FL setting. Theory and experiments accompany in the pre-print.
Joined Meta (Facebook) as a Research Scientist.
Passed my PhD preliminary exam! Onto the final stretch!
Started working as a Graduate Research Assistant for Prof. Xiaojun Lin. We are continuing to focus on developing theoretical results that guarantee exact basis recovery under privacy for sparse models in the under-determined domain.
Accepted into the OpenMined Padawan Open Source program! Working with Ishan Mishra (Engineering Tech Lead) and Ionesio (Team Lead) for integrating into the core open-source team of PySyft. PySyft is a data science library that enables machine learning without transferring data from the client.
Our recent pre-print on optimal client sampling for differentially private federated networks (LOCKS: User Differentially Private and Federated Optimal Client Sampling) is now available on Arxiv.
Updated blog with most recommended workflow tools.
Time for some open-source. Started contributing to Hugging Face's Gradio and Diffusers libraries. Working with The Anvil (at Purdue) to generate team-matching algorithms using Natural Language algorithms.
Completed my internship at Meta working on Federated Semi-Supervised Learning (vision) algorithms.
Excited to be a part of the Cohere for AI Initiative led by Sara Hooker!
Our papers 'Private Hypothesis Testing for Social Sciences' and 'PowerGraph: Using neural networks and principal components to multivariate statistical power trade-offs' have been accepted (poster and talk) to the ICML workshops of Theory of Differential Privacy and AI for Science held alongside ICML 2022 in Baltimore, USA.
Excited to be back at Meta for a summer internship!
Our paper 'Private Hypothesis Testing for Social Sciences' is now available on Arxiv.
I launched the 101 Days of NLP — a popular way of entering a new field. Over the next few months I will be studying NLP in every form that I can. You can follow along on the curious page. All released code will be 100% open-source and reusable.
We presented our preliminary work, 'How to promote open science under privacy,' an article at the confluence of Social Sciences and Privacy at the Psychological Sciences Department at Purdue University. We are working on a tool in R/Python to enable easy sharing of datasets under privacy guarantees.
Our article 'PowerGraph: Using neural networks and principal components to determine multivariate statistical power trade-offs' has been accepted for an Individual Oral Presentation at the International Meeting of the Psychometric Society (IMPS) 2022.
We released a preprint of our recent work on efficiently graphing multivariate statistical power manifolds with supervised learning techniques. Manuscript now available at PowerGraph: Using neural networks and principal components to multivariate statistical power trade-offs.
We discuss recent progress in graphing multivariate statistical power manifolds with novel Machine Learning at the MCP Colloquium at Purdue University. Work done with the SuperPower group. Slides and paper are coming soon.
Our article Towards Quantifying the Carbon Emissions of Differentially Private Machine Learning was accepted at the ICML Workshop on Socially Responsible Machine Learning.
Started internship as PhD SWE at Facebook, Menlo Park.
Started a new blog on differential privacy and federated learning along with my journey in my PhD.
Added a new blog post for workflows in Machine Learning Systems at ML Workflows for Research Scientists. Published in Editor's Picks on Towards Data Science publication (at link).
Promoted to Machine Learning Team Lead in the SuperPower research group.
Our article FedPerf: A Practitioners' Guide to Performance of Federated Learning Algorithms, got accepted to NeurIPS 2020 Preregistration Workshop. We were among the 8 papers invited for a contributed talk for the workshop. Further open-source updates coming soon!
Submitted research article FedPerf: A Practitioners' Guide to Performance of Federated Learning Algorithms to NIPS 2020 Preregistration Workshop.
Joined SuperPower Team (PIs: Dr. Erin Hennes and Dr. Sean Lane) as a Graduate Research Assistant, designing automated and smart algorithms for parameter space exploration using Machine Learning techniques.
Joined OpenMined as a Research Scientist.
Published my first blog.
Started the website. My research paper DFC: Dynamic UL-DL Frame Configuration Mechanism for Improving Channel Access in eLAA, (at NeWS Lab at IIT Hyderabad), was published in IEEE Networking Letters.