Research Engineer, Polaris, DeepMind

About the job

From creating experiments and prototyping implementations to designing new architectures, engineers work on real-world problems including artificial intelligence, data mining, natural language processing, hardware and software performance analysis, improving compilers for mobile platforms, as well as core search and much more. But you stay connected to your research roots as an active contributor to the wider research community by partnering with universities and publishing papers.

Engineers on the Polaris team design tasks that test Gemini’s ability to complete software engineering work. A task is essentially an in-depth examination that evaluates how well Gemini performs in a specific, complex scenario. The primary components of a task are: a prompt that instructs Gemini what to do and a grader or rubric that assigns a numerical score to the model’s attempted solution.

Our team, Polaris, is distinguished primarily by how much manual effort we put into building individual tasks. Rather than generating a large number of tasks using a synthetic data pipeline, engineers on the Polaris team prioritize creating a smaller number of tasks that more closely capture the realism and complexity of real-world software engineering. A single well-designed task typically takes a Polaris engineer roughly one week to complete, from the initial ideation stage to final approval.

Minimum qualifications:

• Experience in software development or programming.
• Experience using AI-assisted coding tools, developer agents, or LLM-based workflows for software development.
• Experience using coding agents for performing work.

Preferred qualifications:

• Experience participating in competitive programming, mathematics competitions (e.g., IMO, USAMO, Putnam, or IOI), or hackathons.
• Experience designing rubrics, evaluation frameworks, or benchmark datasets for machine learning models.
• Demonstrated ability to quickly adapt to and learn new programming paradigms, tools, and technical stacks.

Responsibilities

• Create tasks that will be used to measure Gemini’s ability to perform long-horizon software engineering tasks, and launch entirely new projects within DeepMind aimed at improving Gemini’s training data and evaluations (evals).