AI Research Engineer, Data Pipelines

Engineering

Full time

Remote · Latin America

Revelo is the data lab specializing in code.

Since 2023, Revelo has been building RL environments, evals and specialized mid- to post-training data for leading AI labs. We help them train and evaluate their models on everything LLMs do through code: agentic workflows, software engineering, cybersecurity, terminal work, web apps, games, 3D graphics, infrastructure, ML engineering and scientific computing. Our tasks draw on human expertise from a network of 400,000+ engineers that we have vetted for more than ten years, scaling alongside robust AI engineering pipelines.

We are growing our delivery team

Our research team explores new kinds of training data and proves them on samples. When an AI lab commits to a dataset, the delivery team builds it: we design the pipeline that produces it, run it at the scale of thousands of tasks, and own the quality of every task we ship.

As an AI Research Engineer on the delivery team, you will take what frontier research has learned about RL environments, verifiers, evals and synthetic data, and make it work on real projects for AI labs. The focus is less on publishing original research and more on turning the best of it into pipelines that ship. This is an individual-contributor role.

Apply if you want to collaborate directly with AI research teams from Anthropic · Google DeepMind · Meta · Amazon · Mistral · Tencent · ByteDance

Core responsibilities

Turn research into working pipelines

  • Read recent research on RL environments, verifiers, LLM judges, synthetic data and reward hacking, and decide what holds up on a real project. Turn it into pipeline steps that run every day.
  • Take a lab's data request from a working sample to a production plan: what a good task looks like, how it is verified, how its difficulty is measured against frontier models, and what each task costs.
  • Test ideas on small batches with clear metrics before spending at scale, and bring what you learn in production back to the research team.

Deliver at scale, with quality you can show

  • Own the quality of a delivery end to end: verifiers, automated review and repair, judge calibration, and spot checks by human experts.
  • Watch the numbers that decide a project: how many tasks survive each step, the cost per accepted task, how often frontier models pass, and why tasks fail. Find where the pipeline leaks and fix it.
  • Work with lab research teams through each round of feedback, and analyze where and why models fail on the tasks you shipped, so each delivery makes the next one better.

Qualifications

Required

  • Hands-on AI engineering experience building and running LLM-based systems: agents, tool use, evals, and debugging model behavior.
  • Ability to read research and turn it into working code within short delivery cycles, with judgment about what transfers to a real project.
  • Experience shipping software projects (code, data, benchmarks) at volume, with systematic quality checks.
  • Sound experimental judgment applied to production: choosing baselines and metrics, running small tests before large runs, and producing reproducible results.
  • Comfort with ambiguity. There are many ways to build the same pipeline: you choose one, validate it with quick tests, and keep quality steady as volume grows.
  • Clear technical writing and effective collaboration with researchers, engineers and lab teams to resolve ambiguous requirements.

Preferred

  • Work on production data for code generation, software engineering, scientific computing or multimodal tasks.
  • Experience with LLM post-training, RL environments or verifiers, including reward hacking and robustness testing.
  • Academic research experience (for example a master's or PhD project or a research assistantship), publications, or open-source contributions.

Cool stuff the delivery team is doing

Agent pipelines at scale

Pipelines of agents that generate, verify, repair and calibrate thousands of RL tasks, with yield and cost measured at every step.

Researchers in the loop

Scientists tell us where AI fails in their field. We turn each interview into many tasks, and they review what we built.

Why models fail

Read agent trajectories across model families to learn why frontier models fail, and turn what we find into better data.

Judges from several AI companies

Quality checks graded by panels of models from different companies, so no single model's blind spots decide what ships.

Multimodal projects

Training data across video, audio and code, taken from pilot to full delivery.

Benchmarks become datasets

When one of our benchmarks shows a gap labs care about, we turn it into a dataset they can train on.

Apply for this role

Job ID

503

Type this number in the last question of the form.