Applied Research — RL & Agents

Human Intuition Inc.; New York (NY); On-site; USD 120,000 - 180,000; Full time

Job summary

Human Intuition Inc. is building an autonomous company that turns institutional knowledge into software capable of learning from judgments and outcomes. Teams aim to deploy agents that act within defined constraints and improve through feedback.

The role focuses on researching how agents acquire judgment to perform multi-step tasks, linking learning methods to real business workflows and evaluating progress via trustworthy experiments.

Qualifications

• Strong machine learning fundamentals and hands-on experience with model training or post-training.
• Strong Python and practical familiarity with a modern deep learning framework.
• The ability to design experiments, debug learning systems, and reason carefully from incomplete results.
• Evidence of original technical work through research, open-source software, or deployed systems.
• Clear technical writing and an interest in understanding the business tasks behind an experiment.

Responsibilities

• Develop task environments and agent harnesses that capture meaningful decisions and consequences.
• Investigate supervised fine-tuning, reinforcement learning, and other post-training approaches suited to the available feedback.
• Design rewards and training curricula while testing for shortcuts, reward exploitation, and poor generalization.
• Run controlled experiments on tool use, memory, planning, recovery, and long-horizon task completion.
• Work with the evaluations team to separate learning gains from benchmark leakage or changes in the surrounding system.
• Partner with engineers to make successful research reproducible and useful in deployed agents.

Skills

Strong ML fundamentals

Python proficiency

Experiment design

Original technical work

Technical writing

Job description

Building the autonomous company

Human Intuition is building the autonomous company. Businesses run on accumulated judgment: how to interpret a situation, choose an action, and learn from its consequences. Much of that knowledge lives in people, even when the decisions they make leave traces in software.

We are working to make that judgment learnable. A business has defined systems, tools, permissions, histories, and objectives. Those boundaries create an opportunity to build agents that learn from how work is done, act within clear constraints, and improve through feedback. Our ambition is to turn the knowledge inside institutions into software that compounds.

The role

Research how agents can acquire the judgment needed to carry out consequential, multi-step work. You will connect learning methods to concrete business tasks, studying how policies improve from demonstrations, verifiable outcomes, and feedback. The goal is reliable progress on real workflows, measured through experiments the team can trust.

What you’ll do

• Develop task environments and agent harnesses that capture meaningful decisions and consequences.

• Investigate supervised fine-tuning, reinforcement learning, and other post-training approaches suited to the available feedback.

• Design rewards and training curricula while testing for shortcuts, reward exploitation, and poor generalization.

• Run controlled experiments on tool use, memory, planning, recovery, and long-horizon task completion.

• Work with the evaluations team to separate learning gains from benchmark leakage or changes in the surrounding system.

• Partner with engineers to make successful research reproducible and useful in deployed agents.

What you’ll bring

• Strong machine learning fundamentals and hands-on experience with model training or post-training.

• Strong Python and practical familiarity with a modern deep learning framework.

• The ability to design experiments, debug learning systems, and reason carefully from incomplete results.

• Evidence of original technical work through research, open-source software, or deployed systems.

• Clear technical writing and an interest in understanding the business tasks behind an experiment.

Useful experience

Reinforcement learning, imitation learning, reward modeling, agent memory, distributed training, or learning from operational traces. We value a clear account of what you tried, what failed, and what you learned.

What success looks like

An improvement survives rigorous evaluation, transfers to representative workflows, and can be reproduced by the rest of the team.