ML Engineer
U.S. Department of Defense
AI safety for machine learning systems in defense contexts — evaluating robustness, failure modes, and safe deployment practices for high-stakes applications.
Anirudh Venkatapuram
ML Engineer
U.S. Pentagon
About Me
I am an ML engineer at the Pentagon, where I work on AI safety — evaluating how models fail, where they break under pressure, and what it takes to deploy them responsibly in high-stakes environments. My interests sit at the intersection of technical AI safety and real-world governance: fairness, privacy, robustness, and the institutions needed to navigate catastrophic risk from advanced AI.
I graduated from the University of Virginia with a double major in computer science and mathematics (probability & statistics), where I was a Rodman Scholar. I led the United Nations AI Hub for AI for Good, co-founded Machine Learning @ UVA, and conducted research on fairness–privacy trade-offs in multimodal AI with mentors at the Microsoft AI for Good Lab and Prof. Ferdinando Fioretto's group at UVA.
I'm interested in building organizations and tools that make AI systems safer before they reach the world.
Experience
External AI Consultant
Johns Hopkins Applied Physics Laboratory
Analyzed COLMAP-based 3D reconstruction pipelines; built automated benchmarking for NeRF and 3DGS models across drone, satellite, and smartphone imagery.
AI Hub Leader
United Nations · AI for Good
Led development of multi-modal active learning benchmarks measuring energy efficiency per accuracy point; connected privacy budgets and bias propagation to AI governance standards including GDPR.
Digital Experience Team
Adobe Inc
Built and deployed a RAG chatbot on GPT-4 with a hallucination detection layer that reduced bad responses by over 60%; indexed 200+ documents with ChromaDB and OpenAI embeddings.
Publications
Contrastive Activation Steering for Demographic Bias Mitigation in Code Language Models; The Responsibility Gap in AI: Fairness Metrics and Privacy Tools as Sociotechnical Infrastructures
Dual technical and STS capstone investigating demographic bias in code language models via contrastive activation steering, alongside an analysis of how fairness metrics and privacy tools function as sociotechnical infrastructures in AI governance.
Projects
Fairness–Privacy Trade-offs in LoRA
Designed parameter-efficient fine-tuning strategies to evaluate how fairness, privacy, and accuracy trade off in multimodal AI systems. Mentored by researchers at Microsoft AI for Good Lab and UVA.
Machine Learning @ UVA
Co-founded UVA's ML community and led a team building modular RAG benchmarking pipelines. Produced technical evaluations of retrieval architectures including RAPTOR and LightRAG.
UN Active Learning Benchmarks
Built benchmarks across ResNet18 and DistilBERT measuring energy cost per accuracy point gained, surfacing how privacy and bias interact with model efficiency and regulatory compliance.
Streaming Predictive Maintenance
Real-time pipeline for UVA Sustainability integrating Kafka, Spark Structured Streaming, and AWS Lambda. Delivered guidelines for privacy-preserving, sustainable analytics.
3D Reconstruction Benchmarking
Automated evaluation framework for NeRF and 3D Gaussian Splatting models at JHU APL, stress-testing robustness across heterogeneous real-world imagery sources.
Responsible RAG at Adobe
End-to-end RAG system with rule-based graph constraints for hallucination reduction, deployed on AWS with PostgreSQL-backed evaluation metrics and a React frontend.