Philipp Schmid is a staff engineer in developer experience and developer relations at Google DeepMind. He writes worked guides to adapting models, deploying them and building agents around them.[4]
Start with How to Fine-Tune LLMs in 2024 with Hugging Face →
Links
X 
GitHub Blog Staff Engineer, Developer Experience and Developer Relations · Google DeepMind
Career
Contributions
Teaching model deployment
Schmid’s public work includes a Hugging Face and SageMaker workshop series and deep-learning notebooks. These materials turn model training and deployment into examples developers can run and adapt. His profile also lists clipper.js, a tool for converting web pages to Markdown.[1]
A complete fine-tuning workflow
Schmid’s 2024 fine-tuning guide follows a concrete text-to-SQL use case: define the task, prepare example conversations, train an open model with Hugging Face’s TRL library, evaluate it and deploy it. Fine-tuning changes a model’s weights using task examples, so the useful lesson is the whole data-to-evaluation loop rather than a single training command. His later guide adds distributed training and optimization, while directing newcomers back to the earlier walkthrough.[5]
Keeping agent secrets outside the sandbox
His September 2026 credentials tutorial explains how a managed agent can call an external service without exposing the raw API key to code running in its Linux sandbox. The key remains on a server; a proxy inserts it into requests for approved domains. The example uses a GitHub review workflow to show credential registration, binding and use. This is documentation and example code for a team-built API, not evidence that Schmid alone designed the service.[6]
Ideas
Schmid also writes about the harness, the software that provides a model with tools and manages its execution. In his Bash essay, he reports simplifying a personal harness to a shell interface and a media viewer. His argument is that a shell lets a model compose existing programs without requiring a separate tool for every file operation. He keeps specialized interfaces where they add useful capabilities, and describes isolation, diagnostics and output limits as responsibilities of the surrounding harness.[7]
Tools, writing & talks
Writing, research & talks
Start here
How to Fine-Tune LLMs in 2024 with Hugging Facephilschmid.deA text-to-SQL walkthrough covering dataset preparation, supervised fine-tuning, evaluation and deployment.
Begin with this end-to-end example to learn where data preparation and evaluation fit around training. It assumes Python and access to a suitable GPU; the later guides build on that foundation.
Author[5]Google DeepMind's Philipp Schmid presents the Gemini Interactions API, which replaces user/model turns with steps, and managed sandboxes where the Antigravity agent runs code and shares environments, loading skills and AGENTS.md automatically.
Catalog workOriginal source- Credentials API for Gemini Managed Agentsphilschmid.de
A worked example keeping service credentials outside code executed by an agent.
Author[6] - How Foundational Models Became Superhuman in Bashphilschmid.de
An account of simplifying an agent harness around shell workflows and explicit execution controls.
Author[7]