Patrick Löber is a member of the technical staff and developer relations engineer at Google DeepMind, working on the Gemini API, Google AI Studio, and Gemma. As the creator of Python Engineer, he helps developers understand machine learning through working implementations and build practical applications with increasingly capable multimodal models.
Löber earned a master’s degree in medical engineering at Friedrich-Alexander University Erlangen-Nuremberg, specializing in software engineering, computer vision, and machine learning. His graduate research applied machine learning to identifying vascular occlusions in stroke patients.
After working as a software engineer, he developed Python Engineer into a collection of tutorials, courses, and open-source learning projects. MLfromscratch implements algorithms including regression, decision trees, support-vector machines, and principal component analysis with NumPy; pytorchTutorial teaches practical deep learning; and snake-ai-pytorch demonstrates reinforcement learning by training an agent to play Snake.
In February 2022, Löber joined AssemblyAI as a developer advocate, creating technical tutorials, demonstration applications, workshops, and an online creator community around speech technology. He joined Google DeepMind in February 2025, where his work includes documentation, SDK examples, developer workshops, and launches involving Gemini, Nano Banana, and Veo.
Machine learning through implementation. Löber’s educational repositories make algorithms inspectable: learners implement mathematical operations, training loops, environments, and agent behavior themselves instead of encountering models exclusively through high-level abstractions.
Any-to-any multimodal agents. His multimodal agent architecture distinguishes understanding several input formats from generating several output formats. Gemini reasons across documents, images, audio, and video, then invokes specialized image-generation and speech-generation models through function calls. A NotebookLM-style educational assistant can consequently decide whether a concept needs a diagram, spoken explanation, or additional supporting material.
Agentic orchestration and developer infrastructure. Löber designs adaptive tool-calling loops that select useful outputs and assess whether more material is needed. His implementation guidance addresses larger file uploads, targeted analysis of long recordings, and context caching; a public Gemini demonstration combines Google Maps and Search tools to connect model reasoning with external services.
Multimodal retrieval and local models. His Gemini and AI Studio workshop covers multimodal search, real-time interaction, tool-enabled coding assistants, and local Gemma agents, extending the same practical teaching approach across hosted and locally runnable systems.
▶ Watch ↗