Hands-on experience with LLM fine-tuning and reinforcement learning.
Strong Python and PyTorch programming skills.
Capable of navigating research uncertainty and production limits.
Familiar with environment design and evaluation metrics.
3-5 years of relevant industry experience preferred.
Responsibilities
Design and run model improvement workflows for AI agents.
Create and evaluate environments that challenge static benchmarks.
Build high-fidelity data trajectories for model training.
Develop robust reward and evaluation systems for model behavior.
Manage experiments across distributed GPU clusters and ML infrastructure.
Translate customer research goals into practical projects.
Benefits
Collaborative work with leading AI researchers and labs.
Opportunity to own entire projects from concept to execution.
Dynamic role with exposure to cutting-edge technologies.
Join a team with expertise from top-tier tech companies.
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.
