About the role

We are building a recursive self-improvement system — a machine learning system that iteratively improves itself through feedback, evaluation, and automated learning loops. You will help build the engineering pipeline that keeps these loops fast, reliable, and trustworthy: the training and evaluation pipelines, the reward and feedback signals, and the safeguards that prevent a self-improving system from silently degrading or gaming its objectives.

This is an engineering-first role with deep reinforcement learning requirements. You should be equally comfortable writing robust production ML code and reasoning about reward design, credit assignment, and why feedback-driven systems become unstable.

What you'll do

• Build and maintain the training, evaluation, and deployment loops at the core of the self-improvement system, with a strong emphasis on reproducibility and reliability.

• Design and implement reward and feedback signals; investigate and mitigate reward hacking, specification gaming, and distribution drift.

• Build evaluation harnesses and metrics before models — because a self-improving system is only as safe as its measurement of “better.”

• Own data pipelines and automated data flywheels that feed the learning loop.

• Debug subtle model-quality regressions and stabilize training and feedback loops that go non-stationary.

• Collaborate with research and product to turn methods into robust, shippable systems.

What we're looking for

Must-Haves:

Strong MLE fundamentals (non-negotiable)

• Excellent Python and clean, well-tested ML training code.

• Solid grasp of data pipelines, distributed / large-scale training, and experiment tracking.

• The instinct and skill to debug why a model silently got worse — not just why it crashed.

Hands-on experience with a feedback or learning loop (at least one).

• Built or owned part of a feedback loop — a reward model, an evaluation harness, or the data pipeline for an RLHF/RLAIF or active-learning system.

• Ran a retraining or continual-learning pipeline where a model consumed its own predictions or production data (e.g. ranking, recommendations, fraud, spam).

• Fine-tuned LLMs with human or AI feedback, or built agentic evaluation harnesses.

Reinforcement learning foundations and curiosity.

• Working knowledge of reward modeling, on-policy vs. off-policy tradeoffs, and credit assignment (does not need to be a research-level RL expert).

• Has seen — or can reason clearly about — feedback-system failure modes: reward hacking, specification gaming, feedback loops amplifying errors.

• Comfortable evaluating non-stationary systems (systems whose behavior and data distribution change over time).

Systems and evaluation instinct.

• Builds the eval before the model; treats measurement as a first-class deliverable.

• Has shipped an ML system into production and kept it healthy over time.

Strongest signal

Bonus, not required:

The ideal candidate has built or shipped a full system that improved from its own outputs or feedback end to end. This is rare at this level, so treat it as a standout differentiator rather than a filter. Examples:

• RLHF / RLAIF pipelines

• Self-play systems

• Active-learning loops

• Automated data flywheels

• Agentic evaluation harnesses

Nice to Have:

• PhD or MS in RL / ML paired with real production experience (either the science or the engineering half alone is fine if the other is strong).

• Experience at a lab or company doing RLHF, agents, or large-scale ML infrastructure.

• Familiarity with LLM fine-tuning, evaluation frameworks, or agent orchestration.