Ajinkya Kiran Mulay

ML Research Scientist · AI Safety & Interpretability

Research Interests

Frontier AI Safety, Interpretability, and Evaluations: I focus on mechanistic understanding of neural networks and building rigorous evaluation frameworks to detect and control harmful behavior in high-capability systems. During my PhD, I developed theoretically grounded sparsification methods under differential privacy and federated learning constraints, achieving near-SOTA performance with only 8–10% of dense parameters. At Meta Integrity, I design and deploy LLM-driven safety systems at scale across Facebook, Instagram, and Threads.

Experience

Meta Platforms, IncMenlo Park, CA
  • Own safety evaluations for integrity LLMs, building automated red-teaming/robustness suites (prompt-injection variants, policy-bypass probes, adversarial scenarios) to prevent regressions pre-rollout.
  • Develop user-experience metrics and operationalize them via stratified evaluation and holdouts to understand user retention patterns.
  • Translate new attacker patterns into targeted RL evals and related mitigations, closing the loop from incident signal spikes to launch decisions.
  • Stack: PyTorch, Evals, AI safety, LLM tooling, Red teaming.
Meta Platforms, IncMenlo Park, CA
  • Designed and deployed LLM systems for large-scale spam/scam detection across Facebook, Instagram, and Threads, achieving 4x TPR improvements with detection latency under 4 hours.
  • Fine-tuned large (>70B) language models with retrieval-augmented safety workflows and production instrumentation to detect spam anomalies.
  • Built large-scale behavior monitoring with LLM-assisted labeling and DBSCAN-style clustering to surface coordinated/novel attacker behaviors; curated safety-critical datasets (>1M samples) for post-training and fine-tuning.
  • Stack: PyTorch, PHP, Integrity, Spam.
ECE, Purdue UniversityWest Lafayette, IN
  • Authored SPriFed-OMP, a differentially private federated learning algorithm for sparse basis recovery in high-dimensional regimes (p ≫ n). Achieved accurate recovery with n = O(√p) samples using only 8–10% of model parameters.
  • Combined OMP-style coordinate selection with differential privacy noise injection, isolating informative parameter subsets aligned with interpretable functional components.
  • Stack: PyTorch, Differential Privacy, Federated Learning, Sparse Learning.
Meta (Facebook)Menlo Park, CA
  • Designed and deployed a modular end-to-end production stack for Federated Semi-Supervised Learning (FSSL) vision tasks.
  • Replicated performance benchmarks with FixMatch and SimCLR on real devices.
  • Stack: C++, TorchScript, Python, PyTorch.
Meta (Facebook)Menlo Park, CA
  • Developed a fast, scalable private ML algorithm using PCA with differential privacy, outperforming SOTA by 15% (test accuracy).
  • Improved performance-to-privacy trade-off by more than 35% via varying tree restarts for DP-FTRL.
  • Stack: Python, PyTorch, Differential Privacy, Federated Learning.
SuperPower Research, Psychological Sciences, PurdueWest Lafayette, IN
  • Designed a statistical-power modeling engine (NIH-funded) achieving <5% error while reducing computation by 90% vs. SOTA.
  • Proposed semi-supervised data augmentation and dimensionality reduction methods improving engine stability and power estimation variance.
  • Stack: PyTorch, R, Bayesian Learning, Hypothesis Testing, Differential Privacy.

Other Research Experience

AI Safety CampRemote
  • Collaborate with a cross-disciplinary group modeling frontier AI risk, focusing on how "AI slop" (high-volume, low-signal synthetic content) is created.
  • Design and run evaluation frameworks to detect and quantify AI slop across platforms and modalities.
  • Conduct interpretability analyses on how slop-like data distributions affect internal model representations (neuron/feature activations).
BlueDot Impact Technical AI Safety CourseRemote
  • Completed structured coursework on technical and governance dimensions of AI safety, covering alignment, interpretability, and catastrophic risk from frontier systems.
OpenMinedRemote
  • Explored the relationship between Differential Privacy and Adversarial Robustness; quantified DP/FL impact on real-world systems (FedPerf).
  • Stack: PyTorch, PySyft, Git.

Education

Purdue UniversityWest Lafayette, IN
  • Advised by Prof. Xiaojun Lin  |  GPA: 3.6/4.0
  • Thesis: Developed private and non-private sparse learning algorithms with provable convergence under extreme sample scarcity, achieving near-SOTA accuracy with ≤10% of dense parameters.
IIT HyderabadHyderabad, India
  • Advised by Prof. Bheemarjuna Reddy  |  GPA: 8.88/10
  • Research Focus: Inference-aware game-theoretic framework for unlicensed LTE and Wi-Fi bands.

Skills

Frontier LLM SystemsFine-tuning ≥70B models, RAG-based safety workflows, alignment evaluation.
Efficient MLSparse training, parameter reduction, structured optimization, extreme low-sample regimes (n ~ √p).
Safety & IntegrityLarge-scale spam/scam detection, high-TPR safety filters, behavior anomaly detection.
TheoryConvergence proofs, DP-SGD, private PCA, sparse recovery, rational iteration for optimizer orthogonalization.
Kernel EngineeringTriton, block-sparse FlashAttention, CSR-format sparsity masks, H100 kernel optimization.
TechnicalPyTorch, Triton, C++, Golang, TorchScript, PHP, distributed ML pipelines.

Projects

Halley-Gram-Muon: Cubic Convergence for Optimizer OrthogonalizationRemote
  • Replaced Newton-Schulz polynomial in GramMuon (used in Kimi K2, GLM-5) with a Halley rational iteration in Gram space achieving cubic convergence. Pareto-dominant over GramMuon T=5 on WikiText-103 and enwiki8: better model quality and 16–29% faster wall-clock.
  • Stack: PyTorch, Triton, language model training.
Activation Sparsity as a Scheming Signal — Apart Research AI Control HackathonRemote
  • Mechanistic, output-independent scheming monitor: scheming agents produce more uniform MLP activations (lower Gini) across 15/18 layers. Threshold detector achieves AUROC = 0.745 with no output access (p = 0.0003, Cohen's d = 0.701), consistent across Qwen2.5 1.5B and 3B.
  • Stack: PyTorch, Qwen2.5, mechanistic interpretability, activation probing.
Block-Sparse Causal Attention Kernel — Paradigm Attention Kernel ChallengeRemote
  • Triton kernel for block-sparse causal attention on H100 (CSR mask, head_dim=128). Applied OMP column-reuse intuition: KV-block reuse tiling across query row groups sharing support in sliding-window/banded patterns.
  • Stack: Triton, PyTorch, H100, block-sparse FlashAttention.
Sparse Structured Agent for Financial Document QA — Sentient Arena OfficeQARemote
  • Agentic reasoning system for multi-hop QA over Treasury Bulletin corpora, grounded in Sanskrit Kāraka grammar: query roles (agent, object, instrument) map to structured retrieval steps, with OMP-based decomposition identifying the minimal supporting document set.
  • Stack: Goose, Python, OMP-style retrieval, agentic reasoning, Sanskrit Karaka grammar.

Journal Publications

Ajinkya Kiran Mulay, Xiaojun Lin. "SPriFed-OMP: A Differentially Private Federated Learning Algorithm for Sparse Basis Recovery." Transactions of Machine Learning Research, Purdue University.
Ajinkya Kiran Mulay, Anand Basawade, Bheemarjuna Tamma, Anthony Franklin. "DFC: Dynamic UL-DL Frame Configuration for Improving Channel Access in eLAA." IEEE Networking Letters, IIT Hyderabad.

Early Research Work

Ajinkya Kiran Mulay, Xiaojun Lin. "Humming-Bird: A Forward-Backward Based Differentially Private Federated Learning Algorithm for Sparse Basis Recovery." In Review, Purdue.
Ajinkya Kiran Mulay, Xiaojun Lin. "Humming-Bird+: Batched and General Differentially Private FL Algorithms for Sparse Basis Recovery." In Preparation, Purdue.

Workshop Presentations

Ajinkya Kiran Mulay, Sean Lane, Erin Hennes. "PowerGraph: Using neural networks and principal components to multivariate statistical power trade-offs." AI for Science Workshop, ICML, July 2022 (Non-Archived). SuperPower Lab, Purdue.
Ajinkya Kiran Mulay, Sean Lane, Erin Hennes. "Private Hypothesis Testing for Social Sciences." Theory and Practice of Differential Privacy Workshop, ICML, July 2022 (Non-Archived). SuperPower Lab, Purdue.
Rakshit Naidu, Harshita Diddee, Ajinkya Kiran Mulay, Aleti Vardhan, Krithika Ramesh, Ahmed Zamzam. "Towards Quantifying the Carbon Emissions of Differentially Private Machine Learning." Socially Responsible ML Workshop, ICML, July 2021 (Non-Archived). OpenMined.
Ajinkya Kiran Mulay, Tushar Semwal, Ayush Agrawal. "FedPerf: A Practitioners' Guide to Performance of Federated Learning Algorithms." Pre-Registration Experiment Workshop, NeurIPS, December 2020 (Archived). OpenMined.

Honors & Awards

2026Third Prize, Sentient Arena Cohort 0 OfficeQA Leaderboard.
2025Winner, Grokipedia track, XAI Hackathon, CA, USA.
2023Graduate Research Assistantship, ECE Department, Purdue.
2020Graduate Research Assistantship, SuperPower Group, Purdue.
2018Winner and World Finalist, Microsoft Imagine Cup Japan National Final (Emergensor Startup).
2018Winner, Third Business Plan Competition, University of Tokyo.
2017Two-Year Graduate Teaching Assistantship, ECE Department, Purdue.
2017India-Japan Engineering Program Research Scholarship, University of Tokyo.
2016Undergraduate Teaching Assistantship, IIT Hyderabad.
2016Special Recognition and 8th Rank for Young Team, IEEE Signal Processing Cup.
2014Academic Excellence Award, IIT Hyderabad.
2010National Talent Search Examination (NTSE) Scholar, Govt. of India.

Invited Talks

2023Using neural networks and principal components to optimize multivariate statistical power trade-offs. Modern Modeling Methods Conference (Accepted; unable to attend).
2023Privacy of Noisy SGD. ML Theory, Cohere for AI.
2022How to promote open science under privacy. Psychological Sciences Department, Purdue University.
2022PowerGraph: Using neural networks and principal components to multivariate statistical power trade-offs. International Meeting of the Psychometric Society (Accepted; unable to attend).
2021Graphing multivariate statistical power manifolds with Machine Learning. MCP Colloquium, Purdue University.
2020FedPerf: A Practitioners' Guide to Performance of Federated Learning Algorithms. NeurIPS Pre-Registration Workshop.

Teaching & Mentoring

Jan–May 2022Mentoring Undergraduate Students for the Anvil Co-Founder AI Matching Platform Development.
Aug 2019–May 2020Graduate Teaching Assistant, ECE 27000 — Introduction to Digital Design, Purdue.
Aug 2018–May 2019Graduate Teaching Assistant, ECE 20002 — Electrical Engineering Fundamentals II, Purdue.

Other Services

2026Attendee, EA Global, San Francisco, CA.
2024Reviewer: NeurIPS, ICML TF2M Workshop, ICLR, DMLR, TMLR, ICLR Tiny Papers, Privacy Preserving AI Workshop at AAAI.
2023Reviewer: NeurIPS, AAAI, ICML Tiny Papers, FAccT, ISIT, IJCAI, CHIL. Top Meta-Reviewer, AAAI Representation Learning Workshop.
2022Reviewer: CHIL. Open Source: OpenMined, Gradio by HuggingFace. Professional Grant Reviewer, Purdue.