Career Snapshot

I worked on advanced AI agents and RAG for finance and accounting at Aktus AI. I wrote production code and well comfortable with production engineering pipelines and best engineering practices.  I built and productionized systems for classical and graph RAG (AI agents, multi-agent systems, tool use, embeddings, semantic search, vector databases) and for achieving accurate chart, graph, and long document understanding and multi-step reasoning.

I have also worked in AI safety. I developed RepBend, a novel LLM safety fine-tuning and representation engineering (AI safety and alignment) method to disrupt harmful representations and achieved up to 95% reduction in jailbreak attack success rates. I contributed to multi-modal VLM reasoning and safety benchmarks, alignment frameworks (judge-augmented SFT, on-policy learning). 

I have also been advisor for AIM Intelligence, an AI safety pioneer. 

Previously, I was an AI Engineer at Meta AI and worked on deepfake detection and developed technologies for privacy-preserving machine learning and federated learning. I created, operationalized, and delivered Green Federated Learning, a multi-year million-device work across 5 teams. I was core contributer to Opacus, an open-source library that enables training deep learning models with differential privacy (see blog posts 1 and 2), and FLSim, an open-source library for simulating federated learning systems.

Prior to that, I was a  Visiting Scholar at University of California, Berkeley and member of the Berkeley Artificial Intelligence Research (BAIR). When I was at UC Berkeley, I contributed to Flow, an open-source  deep reinforcement learning-enabled framework for simulation of autonomous and manned cars. 

I earned my PhD degree in Computer Science at the University of Texas at Dallas. I am fortunate to be highly cited (5000 citations). My work was on the intersection of Computer Systems, Edge Computing, and Machine Learning, specifically, on improving quality of service in IoT and deep learning Applications through Fog Computing. I won the UT Dallas Best Dissertation Award.