Mitesh Patel is a senior developer advocacy leader and principal engineer at NVIDIA who helps developers build practical systems for machine learning, generative AI, and robotics. His work combines a research background in robotics and indoor localization and sensor fusion with GPU-accelerated developer workflows for language models and data science.
Patel earned a doctorate in robotics from the University of Technology Sydney in 2014, studying probabilistic models of human activity and assistive robotic systems. He subsequently researched user behavior at Yahoo! Labs before joining Fuji Xerox Palo Alto Laboratory as a senior research scientist. There, he worked on indoor positioning, activity recognition, medical imaging, and systems combining Bluetooth, Wi-Fi, inertial sensors, cameras, and building maps.
His ContextualNet research combined convolutional and recurrent neural networks to infer location from image sequences, evaluating the approach with both a robot and a smartphone. At NVIDIA, he leads a developer advocacy team producing technical workflows, notebooks, and examples that apply GPU acceleration to familiar developer tools. His published work includes accelerated model stacking and hyperparameter optimization and conversational systems for machine-learning tasks.
- HybridRAG for relationship-aware retrieval. Patel combines semantic vector search with knowledge graphs that preserve explicit relationships between entities. His HybridRAG work separates offline document preparation, embeddings, and graph construction from online retrieval, selecting graph, vector, or hybrid approaches according to data structure, relationship complexity, and computational cost.
- Ontology design and triplet quality. He emphasizes that graph-based retrieval depends on accurately extracting entity-relationship-entity triplets from messy documents. Practical improvements include ontology-specific prompts, careful chunk overlap, regular-expression-based cleaning, iterative extraction, and low-rank model adaptation.
- Multihop retrieval within a latency budget. Following several graph connections can uncover relationships that single-hop retrieval misses, but deeper traversal raises production latency. Patel points to GPU-accelerated cuGraph operations and NetworkX integration as ways to investigate richer context while retaining familiar interfaces.
- End-to-end retrieval evaluation. He evaluates both retrieved evidence and generated responses, using external frameworks such as Ragas to examine faithfulness, relevance, precision, and recall. This helps isolate failures in document preparation, retrieval, question interpretation, and response generation.
- Conversational, GPU-accelerated data science. With Allison Ding, Patel coauthored an interactive machine-learning agent that translates plain-language requests into structured operations for data exploration, model training, evaluation, and hyperparameter optimization using GPU-backed tools.
His robotics research has also reemerged through physical AI and simulation: he co-organized a CVPR 2026 physical-AI tutorial and led its robotics-simulation session, covering virtual environments, synthetic data, policy validation, and deployment preparation.
