Research Engineer

Proximal; Hinoba-an; On-site; PHP 1,000,000 - 1,400,000; Full time

Job summary

Proximal is building the research systems needed to identify what models can't yet do, build the tasks required to teach them, measure whether those capabilities improve, and continuously produce the data that frontier models need.

As a research engineer, you will work on open-ended research problems across RL, agents, and data, and share your work with the public through open-source releases, papers and talks.

Qualifications

• Strong technical and engineering abilities; engineering judgment matters far more than formal research experience.
• Creative problem solver with strong experimental instincts; you thrive on ambiguous, open-ended problems.
• Strong research judgment; you can identify promising directions, design rigorous experiments, and draw useful conclusions from messy results.
• Self-directed and relentless; you take ownership of problems end-to-end without heavy process or oversight.
• Execution with excellence, even at high speed. We don't think that shipping fast has to come at the cost of quality and reliability.

Responsibilities

• Design and run experiments to find the gap between frontier models' capabilities and limits across coding tasks.
• Develop methods for automatically generating increasingly difficult tasks as model capabilities improve.
• Design experiments to measure the quality of agent-generated code beyond functional correctness.
• Modify our RL training stack to support end-to-end training with agent harnesses that implement subagents.
• Train specialized models to evaluate and adversarially stress-test data for reward hacking, fairness, and other quality failures.

Skills

Strong engineering judgment

Creative problem solving

Experimental instincts

Self-directed

Open-ended problems

Job description

Proximal is building the research systems needed to identify what models can't yet do, build the tasks required to teach them, measure whether those capabilities improve, and continuously produce the data that frontier models need. We work with frontier AI labs to provide the data and evaluations behind their most capable models.

Our early team has built coding agents and RL infrastructure at companies like Cursor and Prime Intellect, worked at firms like Jane Street, and founded companies that raised millions.

We're growing extremely fast and are backed by top-tier funds including General Catalyst, alongside angel investors from OpenAI, Anthropic, xAI, Meta Superintelligence, Google DeepMind, and Thinking Machines.

About the role

As a research engineer, you will work on open-ended research problems across RL, agents, and data, and share your work with the public through open-source releases, papers and talks. You'll own problems end-to-end, combining research judgment with strong engineering execution.

What you'll do

• Design and run experiments to find the gap between what frontier models can and can't yet do across coding tasks
• Develop methods for automatically generating increasingly difficult tasks as model capabilities improve
• Design experiments to measure the quality of agent-generated code beyond functional correctness
• Modify our RL training stack to support end-to-end training with agent harnesses that implement subagents
• Train specialized models to evaluate and adversarially stress-test data for reward hacking, fairness, and other quality failures

What we look for

• Strong technical and engineering abilities; we find good engineering judgment matters far more than formal research experience
• Creative problem solver with strong experimental instincts, you thrive on ambiguous, open-ended problems
• Strong research judgment; you can identify promising directions, design rigorous experiments, and draw useful conclusions from messy results
• Self-directed and relentless, you take ownership of problems end-to-end without heavy process or oversight
• Execution with excellence, even at high speed. We don't think that shipping fast has to come at the cost of quality and reliability.