Research Engineer, Applied AI

HeyMilo•San Francisco, CA

10h•Onsite

About The Position

HeyMilo is building AI interviewers that automate and improve hiring through conversational AI. We work closely with companies to bring AI into real hiring workflows. We also run an applied AI team that studies where today's models succeed and fail across industries. The Role We're hiring a Research Engineer to join our Applied AI team in San Francisco. You'll design and build reinforcement learning environments with verifiable rewards for specific real-world use cases: the simulators, reward functions, and evaluation harnesses that let us measure and improve how models perform on real work. The work is equal parts ML research and systems engineering. You'll work directly with the founders and our research advisors, taking a use case from problem definition to a reproducible environment that models can be evaluated and trained against.

Requirements

Master's or PhD in AI, Machine Learning, Computer Science, or a closely related field
Solid grounding in reinforcement learning and LLM post-training (reward design, policy optimization, evaluation methodology)
Strong software engineering skills in Python, with code that others can run and build on
Hands-on experience with LLMs: running evaluations, building agentic loops, tool calling, fine-tuning
Comfortable with containers and infrastructure (Docker, Linux, cloud environments) for reproducible experiment setups
Ability to operate in ambiguity and move quickly; comfortable owning a problem end to end
Based in (or willing to relocate to) the San Francisco Bay Area

Nice To Haves

Published research or open-source contributions in ML, RL, or evaluation
Experience with RL/eval frameworks and simulated or sandboxed environments
Experience training or fine-tuning open-weight models at any scale

Responsibilities

Design and build RL environments for specific real-world use cases: realistic simulators, tool interfaces, and episodic task generation with proper isolation and reproducibility
Design verifiable reward functions that score correct intermediate actions as well as end states, and hold up against reward hacking
Build and maintain evaluation harnesses that run task suites across frontier and open-weight models, with clean scoring and cost tracking
Run post-training experiments (e.g. RLVR-style fine-tuning of open models) to validate that your environments produce a learnable signal
Package environments and results for reproducibility, and contribute to research write-ups and published evaluations
Help define which use cases we pursue next, informed by where models are weakest