Role MissionPost-training is the critical bridge between raw model capability and a system that is actually useful, safe, and effective in the real world. You will shape how our agents learn by iterating on training recipes, evaluations, and alignment methods that directly determine what Devin and our future systems can do. This role blends deep research and hands-on engineering. We don't distinguish between the two.
What You'll Accomplish- Post-Training Recipe Development: Iterate on the full stack of datasets, training stages, and hyperparameters that determine model behavior. Measure how choices compound across evals and production performance, not just isolated benchmarks.
- Evaluation Design and Integrity: Build evals that actually capture what matters. The loop never ends: define, optimize, realize the gaps, and rebuild. You'll be responsible for making numbers go up and making sure the numbers mean something.
- Deep Understanding: When training produces results that don't make sense, you dig until you understand why. The goal isn't just to fix it; it's to carry that understanding forward to the next problem.
- Alignment and Agent Behavior: Apply and advance techniques like RLHF, RLAIF, and constitutional approaches to shape how agents reason, act, and collaborate with humans in long-horizon tasks.
- Scaling and Exploration: Measure how performance scales with data and compute, and develop new methodologies when existing ones hit ceilings. We expect both rigor and invention.
Exceptional Candidates Have Demonstrated- A track record of advancing ML systems through post-training, alignment, or related methods: RLHF, RLAIF, preference modeling, reward learning, or equivalent
- Strong fundamentals in probability, statistics, and ML theory. The ability to look at experimental data and distinguish real effects from noise and bugs
- Evidence of original contributions: publications at top venues, open-source impact, or equivalent industry results
- Experience with large-scale distributed training and the debugging that comes with it
- Systems-level thinking: not just model optimization, but understanding how training pipelines, data, and evaluation interact
- Comfort with ambiguity and fast-moving research environments where priorities shift quickly
- We care more about demonstrated capability than credentials. A PhD is one signal among many.
Resources & Environment- Small, highly selective team where research and product move together; prototypes reach real deployment quickly
- Compute is not a constraint: large allocations with training jobs routinely running across thousands of GPUs from day one
- The environment rewards speed, autonomy, and technical depth with minimal process overhead; this is one of the most competitive and fast-moving problems in AI
- Everything needed to operate at frontier scale from day one.
