Software Engineer, Gemini App Skills Platform, DeepMind

Google•Mountain View, CA

3d•$174,000 - $252,000

About The Position

Join DeepMind’s mission to solve intelligence and benefit humanity by building the next generation of agentic AI. As a Software Engineer on the Gemini App Skills Platform, you will contribute to an extensible execution layer that transforms Gemini from a conversational assistant into an action-oriented, universal agent. You will architect systems that allow Gemini to discover, retrieve, and execute modular capabilities (skills) across Google, DeepMind, and third-party ecosystems. The work bridges foundation-model intelligence and real-world execution, including dynamic context engineering, low-latency routing algorithms, and automated evaluation frameworks. The role focuses heavily on LLM behavior and evaluation, with hands-on architectural enhancements to Gemini app production-serving infrastructure, primarily in C++, and Python evaluation pipelines. Your work will help shape the user experience for hundreds of millions of users across web, Android, iOS, and Chrome. DeepMind describes itself as a pioneering AI lab with interdisciplinary teams advancing AI to address global challenges and accelerate product innovation, with safety and ethics as priorities. Individual pay is determined by job-related skills, experience, and relevant education or training.

Requirements

Experience developing software with large language models, agentic workflows, tool use, or retrieval-augmented generation (RAG).
Experience designing evaluation datasets, automated rating systems, or conducting error analysis and prompt optimization for generative models.

Nice To Haves

Experience collaborating cross-functionally across research, product, and platform engineering teams to take novel AI features from prototype to production.
Demonstrated ability to design experimental setups, analyze model telemetry, and define reproducible metrics for complex features.

Responsibilities

Design and build intelligent skill discovery, hybrid routing architectures, and low-latency C++ serving components across Gemini on web, mobile, and Chrome.
Develop dynamic system instructions, prompt representations, and semantic retrieval schemas so Gemini can navigate expanding skill catalogs while avoiding attention dilution and quality regressions.
Design automated evaluation frameworks, including auto-raters, simulated environments, and benchmarking suites, to measure single-turn triggering precision, multi-turn tool use, and parameter validity.
Analyze production telemetry, drive iterative prompt/model hill-climbing to surpass quality thresholds, and establish standards for system reliability and clean architecture.
Partner with DeepMind Research, Search, Vertical teams such as local and shopping, and external Engineering teams to onboard modular capabilities onto the platform.