Masters/PhD-level knowledge of machine learning fundamentals
Hands-on experience with LLM fine-tuning and reinforcement learning
Strong instinct for building agent environments and evaluation harnesses
Fluency in Python and PyTorch, with comfort in lower-level primitives
Strong engineering judgment to navigate research and production constraints
3-5 years of relevant experience preferred, but flexible based on fit
Responsibilities
Design and run model improvement workflows across various training approaches
Create realistic environments and scenarios for agent testing
Inspect and enhance data trajectories for model training
Develop robust reward functions and evaluation systems
Run experiments on distributed GPU clusters and build necessary infrastructure
Translate ambiguous research goals into concrete deliverables for AI labs
Benefits
Collaborative work environment with a focus on impactful projects
Opportunity to work with cutting-edge AI technologies
Ownership of projects from conception to execution
Exposure to a diverse team with extensive industry experience
Flexible work arrangements to support work-life balance
Full Job Description
What You'd Work On
Post-training loops. You'd help design and run model improvement workflows across supervised fine-tuning, preference optimization, and reinforcement learning approaches like GRPO. The work is not just launching training jobs; it's figuring out what signal matters, how to collect it, and whether the model actually improved.
Environment and task design. We build environments that feel real and scenarios that push agents past static benchmark behavior. You'd design tasks, tools, validators, reward signals, and evaluation harnesses that test meaningful capabilities instead of whatever is easiest to measure.
High-fidelity trajectories. You'd create, inspect, and improve the data that teaches models how to behave. That means caring about taste, correctness, edge cases, and whether a trajectory would actually help a frontier model learn.
Reward and evaluation systems. You'd work on reward functions, rubrics, validators, and analysis tools that turn messy model behavior into useful training signal. You should be interested in where evals lie, where rewards get hacked, and how to make measurements more robust.
Training and research infrastructure. You'd run experiments across distributed GPU clusters, work with PyTorch and FSDP, and build the infrastructure needed to support model training, evaluation, and data generation at scale.
Customer research problems. You'd work with frontier AI labs to translate ambiguous research goals into concrete environments, datasets, experiments, and deliverables.
About You
Machine learning fundamentals. You have Masters / PhD-level knowledge of machine learning fundamentals. You understand linear algebra, optimization, stochastic gradient descent, and can reason from first principles when a model or training run behaves unexpectedly.
Hands-on post-training experience. You have worked with or deeply understand LLM fine-tuning, preference data, reward modeling, or reinforcement learning for language models. You know the difference between reproducing a recipe and understanding why it works.
Environment-building instinct. You are excited by agent environments, multi-turn tool use, sandboxed tasks, eval harnesses, and the question of how to test capabilities that do not fit neatly into a benchmark.
Python and PyTorch fluency. You are strong in Python and comfortable building with PyTorch, FastAPI, and modern ML infrastructure. You can work at a high level, but you are also comfortable dropping into lower-level primitives when the abstraction leaks.
Strong engineering judgment. You can move between research ambiguity and production constraints. You know when to iterate quickly, when to be rigorous, and when a result is too fragile to trust.
Experience. No hard rule. Roughly 3-5 years is what we imagine, but more or less experience works if the expectations above resonate with you.
What Makes This Different
The work is concrete. You are not just "touching the latest AI stack." You are building the environments, trajectories, evals, rewards, and training loops that frontier labs use to improve models.
It is research-facing, but production-minded. Our customers are AI researchers and labs pushing the edge of what agents can do. The systems you build need to support real experiments, real users, and real deadlines.
Ownership, not theater. You own whole problems, not isolated tickets. One week you might be designing a new environment; the next you might be debugging a training run, improving a reward function, or scaling an eval pipeline.
Ambiguity is part of the job. There is no fixed playbook for this work. Data, post-training, and agent evaluation are changing quickly. If you need every problem to be fully specified before you begin, this role will be challenging.
The team. Our team is ex-Stripe, Snap, AWS, Microsoft, Airtable - you'll work with a small team who has years of shipping high-impact products over the last decade.
