Founding Data Scientist

We're looking for a Founding Data Scientist to build and post-train language models that power autonomous AI agents. You'll own the complete model improvement lifecycle, from training data and reward design to SFT, reinforcement learning, evaluation, and production deployment. This is a deeply hands-on role working directly with the founders to build proprietary AI models that learn from real-world outcomes.

Responsibilities:

• LLM Post-Training: Fine-tune open-weight models (Llama, Qwen, Gemma) using SFT, DPO, and RL techniques.
• Reinforcement Learning: Design reward functions and optimize model policies using real-world task outcomes and user preferences.
• Agent Data Pipelines: Build high-quality training datasets from agent trajectories, tool interactions, and preference signals.
• Model Evaluation: Develop benchmarks to measure reasoning, tool use, decision-making, and task completion.
• Training and Inference: Run distributed training and high-throughput inference using PyTorch, FSDP, DeepSpeed, vLLM, and SGLang.
• Production Ownership: Deploy, optimize, and continuously improve fine-tuned models.

Requirements:

• Hands-on experience post-training open-weight LLMs using SFT, DPO, and/or RL techniques, with production deployment experience.
• Strong expertise in Python, PyTorch, transformers, attention mechanisms, and LLM architectures.
• Experience with LoRA/PEFT, distributed training (FSDP/DeepSpeed), mixed precision, and multi-GPU infrastructure.
• Practical experience with vLLM/SGLang for high-throughput inference and agent rollouts.
• Experience building LLM training datasets, agent trajectories, preference datasets, and evaluation frameworks.
• Strong understanding of model optimization, experimentation, and failure analysis.
• Good to Have: Experience with PPO, GRPO, RLHF, reward modeling, and long-horizon autonomous agents.

Ideal Candidate:

• Engineers and scientists who have actually trained, fine-tuned, and shipped LLMs, not just built applications using LLM APIs.
• You should enjoy solving complex ML problems, experimenting with model architectures and training techniques, and taking end-to-end ownership in a fast-moving founding team.