ML Research Engineer

Radical AI; New York (NY); On-site; USD 140,000 - 210,000; Full time

Job summary

Radical AI is seeking an engineering-focused researcher to design evaluation datasets, replay historical runs, and ship improvements to our agent's planning and control loops. You will own evaluation and experimentation systems and collaborate with platform engineers.

The profile is engineering-first, ML-strong, and research-capable. You should design data schemas and CI pipelines, and turn ML ideas into controlled experiments that ship improvements with production-quality results.

Qualifications

• 4–8 years building production ML systems, research infrastructure, or data-intensive backends.
• Excellent Python and strong software engineering discipline: testing, typing, packaging, code review, CI
• Direct experience with ML evaluation and experimentation: offline evaluation harnesses, dataset and split design, metric design, and sound reasoning about noisy or underpowered results
• Data pipeline skills: schema design, versioning, orchestration (Airflow, Prefect, Dagster, Ray, or equivalent), object storage
• A reproducibility instinct: pinned environments, seeded runs, versioned artifacts, results that hold up when someone else reruns them six months later
• The ability to take ambiguous system behavior, reduce it to a controlled experiment, and write up what the result does and does not support
• Ability to turn ideas into working prototypes quickly, then harden successful ones into reliable production systems
• Initiative and ownership in an environment where the roadmap is still being written

Responsibilities

• Formulate hypotheses about agent behavior and design controlled experiments, implementing and evaluating improvements to planning, tool use, and decision-making.
• Reproduce historical agent runs from recorded context for fair model comparisons under consistent conditions.
• Develop versioned evaluation datasets, curated task sets, splits, and ground truth that evolve without invalidating past results.
• Build evaluation harness and metrics: decision quality, task success, recovery from error, cost and latency.
• Instrument agent runs, characterize failure modes, and convert recurring failures into regression coverage.
• Prototype, test, and ship control policies for when the agent revises an approach or escalates to a scientist.
• Create experiment infrastructure: data pipelines, artifact and dataset versioning, experiment tracking to enable quick clean comparisons.
• Collaborate with scientists and platform engineers to connect agent behavior to Workcell outcomes.

Skills

Python programming

Production ML systems

CI/testing

Experiment design

Data pipelines

Versioned experiments

LLM agents concepts

Tools

Airflow

Prefect

Dagster

Ray

Langfuse

MLflow

Weigths & Biases

DVC

OpenTelemetry

Prometheus

Grafana

Datadog

Job description

Our scientific agent decides what experiment to run next. This role builds the machinery that lets us understand, evaluate, and continuously improve the agent's decisions.

You will work closely with the Principal Scientist to turn hypotheses about agent behavior into reproducible experiments and production improvements. Your initial focus will be evaluation datasets, replay, trajectory analysis, and the policies governing how the agent plans, responds to results, and recovers from failure.

You will own the evaluation and experimentation systems, contribute research ideas, and carry promising changes from prototype through production in partnership with platform engineers.

The profile is engineering-first, ML-strong, and research-capable. You should be as comfortable designing a data schema and CI pipeline as turning an ML or scientific idea into a controlled experiment and shipping the resulting improvement.

What You'll Work On

• Research and experimentation: work with the Principal Scientist to formulate hypotheses about agent behavior, design controlled experiments, and implement and evaluate improvements to planning, tool use, and decision-making.


• Reproducible replay: reconstruct historical agent runs from recorded context so models, prompts, and policies can be compared under consistent conditions, with clear limits on what those comparisons establish.


• Versioned evaluation datasets: curated task sets, splits, and ground truth that evolve without invalidating past results


• Evaluation harness and metrics: decision quality, task success, recovery from error, cost, and latency, with honest treatment of noise and small sample sizes


• Trajectory analysis: instrument agent runs, characterize failure modes, and convert recurring failures into regression coverage


• Control policies: prototype, test, and ship the logic that decides when the agent revises an approach, pivots, escalates to a scientist, or terminates


• Experiment infrastructure: data pipelines, artifact and dataset versioning, experiment tracking, and the tooling that lets the team run a clean comparison in minutes rather than days


• Close partnership with our scientists and platform engineers to connect agent behavior to what actually happens in the Workcell

What We're Looking For

Required:

• 4-8 years building production ML systems, research infrastructure, or data-intensive backends; strong enough to design, build, and ship end-to-end


• Excellent Python and strong software engineering discipline: testing, typing, packaging, code review, CI


• Direct experience with ML evaluation and experimentation: offline evaluation harnesses, dataset and split design, metric design, and sound reasoning about noisy or underpowered results


• Data pipeline skills: schema design, versioning, orchestration (Airflow, Prefect, Dagster, Ray, or equivalent), object storage


• A reproducibility instinct: pinned environments, seeded runs, versioned artifacts, results that hold up when someone else reruns them six months later


• The ability to take ambiguous system behavior, reduce it to a controlled experiment, and write up what the result does and does not support


• Ability to turn ideas into working prototypes quickly, then harden successful ones into reliable production systems


• Initiative and ownership in an environment where the roadmap is still being written

Nice to have:

• Experience with LLM agents: tool calling, planning loops, context management, trajectory logging, and agent evaluation


• Background in sequential decision making: bandits, Bayesian optimization, reinforcement learning, or off-policy evaluation


• Experiment tracking, agent observability, and artifact versioning tools (Langfuse, MLflow, Weights and Biases, DVC)


• Scientific computing, or experience with materials, chemistry, or instrument data


• Experience operating large-scale inference systems, including GPU scheduling and distributed execution


• Observability experience (OpenTelemetry, Prometheus, Grafana, Datadog)


• Prior work in lab automation, robotics, or a self-driving lab

What Success Looks Like

In your first six to twelve months:

• Representative historical agent runs can be replayed against recorded context and compared with candidate models, prompts, or policies, with clear limits on those comparisons.


• A versioned evaluation suite gates changes to the agent, and the team trusts it enough to act on the result


• Recurring failure modes are documented, categorized, and covered by regression tests.


• Scientific and agent hypotheses can be turned into controlled experiments quickly, with clear conclusions about what the evidence supports and what to try next.


• At least one meaningful agent or control improvement is shipped to production, with measured gains in decision quality or scientific throughput.

Radical AI is an equal opportunity employer. We do not discriminate on the basis of race, color, ancestry, national origin, religion, sex, age, sexual orientation, gender identity and expression, marital status, disability, or veteran status.