🛠 Tech Stack

💼 About This Role

You'll own late-stage training decisions for LLMs, including data mix, annealing schedules, and capability injection. Your work will directly sharpen models used in Devin and Windsurf. This role blends research and engineering in a small, compute-rich environment.

🎯 What You'll Do

• Design and iterate on data mixtures for late-stage training runs.
• Drive targeted improvements in coding, math, and reasoning capabilities.
• Develop synthetic data pipelines for scalable training signal.
• Research and implement context length extension methods.

📋 Requirements

• Deep familiarity with LLM training pipeline end to end.
• Hands-on experience with continual pre-training or annealing for large models.
• Proficiency in Python and deep learning frameworks like PyTorch.
• Strong fundamentals in optimization, statistics, and ML theory.

✨ Nice to Have

• Experience developing synthetic data pipelines for capability improvement.
• Track record of original contributions (publications or open-source impact).
• Comfort operating in ambiguous, fast-moving environments.

🎁 Benefits & Perks

• ⚡ Compute is not a constraint – large GPU allocations from day one.
• 🏢 Small, selective team with rapid deployment of prototypes.
• 🚀 Autonomy and speed rewarded with minimal process overhead.