Senior Machine Learning Engineer, AI Platform & Agentic Apps

The AI Platform & Agentic Apps team builds the agent platform behind every AI agent at Robinhood. Today it gives a growing number of engineers and employees an AI teammate that ships code, queries data, and runs operational workflows on their behalf. We're building toward the same platform powering the agents millions of customers interact with directly, in real time. These agents don't just answer questions — they're designed to take real action across carefully curated meta harnesses. This is agentic AI at real scale, in a regulated financial environment, and it will change how Robinhood works!

As a Staff Machine Learning Engineer on the AI Platform & Agentic Apps team, you will design and build the harness that every agent at Robinhood runs on. A critical part of the role is making those agents trustworthy at scale: trajectory-level evals that measure how an agent reasons and acts, and action guardrails — permission models, approval gates, and sandboxing — built as platform primitives that other teams adopt. You'll be a technical anchor on a growing, high-caliber team, collaborating with product, infrastructure, and fellow ML engineers to take ambitious ideas from zero to one and into production. You'll help define the team's technical direction, mentor engineers, and shape how Robinhood decides an agent is ready to ship. This role offers a rare combination of technical depth, platform-scale impact, and the satisfaction of building systems that genuinely don't exist anywhere else.

* Design and build the core of Robinhood's agent harness — orchestration, tool integrations, context and memory management — so one platform can safely power both high-trust internal agents and tightly scoped customer-facing ones. * Ship agentic applications end to end on that harness, from an ambiguous problem to a production agent that takes real action on behalf of employees or customers, and feed what you learn back into the platform. * Build trajectory-level evaluation systems that score how an agent got to an answer, not just the answer — tool-call correctness, planning and recovery, multi-step task completion — backed by simulation environments and synthetic task generation. * Architect action guardrails as platform primitives: least-privilege tool scoping, permission models, human-approval gates for high-risk or irreversible actions, step and budget limits, sandboxing, and rollback. * Make evals and guardrails products other teams adopt — SDKs, CI regression gates on prompt, model, and tool changes, continuous red-teaming, and production tracing that closes the loop from real traffic back into eval sets and guardrail models. * Set the technical bar through architecture reviews, code reviews, and mentorship, and be the person who can make — and defend with data — the "don't ship" call.