About Thirdlayer
ThirdLayer is solving one of the hardest problems in deploying agents: AI models are generic, but people's work and processes are specific. Our product, Dex, embeds within your computer and team, ingesting a continuous stream of data across your browser and work apps to understand how you actually operate. So when you delegate tasks, agents know exactly the context of where information lives and how decisions get made. We're a deeply technical team with a high bar for talent and a shared belief that exceptional people are the foundation of everything we build.
The role
We're building the training loop that makes agents specific: post-training models on the workflows, context, and feedback our platform captures from real work. As a Research Scientist, you'll own the research questions behind that loop. What signal is actually in a stream of production traces? How do you turn it into environments and rewards a model can learn from? How do you know a model's capabilities actually generalize beyond the eval? You'll design and run experiments at scale, develop new post-training and RL methods, and set the direction for what we train and why. Your work goes into production quickly. The models you train run inside real deployments, and what you learn there feeds the next round of research.
What you will do
- Post-train language models on tasks and environments derived from production data.
- Advance our methods across RL, distillation, and learning from user feedback.
- Design evaluations that tell us whether capabilities generalize beyond the training distribution, and treat evals as a research problem in their own right.
- Build synthetic environments and data pipelines that turn messy production traces into training signal.
- Take large-scale experiments end to end, from hypothesis through analysis to a shipped model.
- Set research direction: decide what's worth training, what's worth measuring, and what to kill.
What we are looking for
- Serious ML research experience, ideally touching post-training, RL, or alignment.
- A visible research track record: publications, open implementations, or models you've trained that people actually use.
- Rigorous experimental design. Your experiments answer the question they were built for, and you don't cut corners to get a result.
- Fluency in Python and a modern ML stack (PyTorch or JAX), including distributed training.
- Comfort working with messy real-world data.
- You can go from paper to prototype to production impact without someone translating for you.
- You want your research running in production.