Role summary
The Company is in a unique position: we develop AI-based products across multiple domains and collaborate with the world’s leading AI companies as partners and customers. This role is directed at measuring improving the security, content safety, and inclusivity of our frontier models. Because we are expanding across multiple pillars of safety, we are looking for specialists with deep expertise in one or more of the following core focus areas
Responsibilities
- Evaluation: Develop datasets and specialized models & algorithms to evaluate/benchmark models & end-to-end systems in our core safety tracks (LLM Security, Agentic Safety, Content Safety, Hallucinations, and ML Fairness).
- Model Pre-Training, Mid-Training, Post-Training: Develop datasets and recipes for filtering training data, developing training datasets & recipes, including components like RL environments and teacher models, across our core safety tracks (LLM Security, Agentic Safety, Content Safety, Hallucinations, and ML Fairness).
- Model & system level techniques beyond post-training: Research & deploy new approaches, like Instruction Hierarchy or Risk Detection.
- Cross-Functional Collaboration: Partner with engineers, data scientists, and research teams across the Company to scale solutions for LLM Security, Agentic Safety, Content Safety, Hallucinations, and ML Fairness.
Requirements
- Master’s or PhD in Computer Science, Electrical Engineering, or a related quantitative field (or equivalent experience).
- 8+ years of proven experience in systems software engineering or machine learning engineering.
- Post-Training Experience: 4+ years of hands-on work experience in post-training of LLMs, including Supervised Fine-Tuning (SFT), Reinforcement Learning (RLHF/RLAIF), safety data generation techniques, ablation studies, and deploying models to production.
- Core Safety Expertise: 1+ years of dedicated experience or research in at least one of the following areas
- LLM Security (backdoors, poisoning, latent behaviors).
- Frontier Risks (deception, manipulation, loss-of-control).
- Agentic Safety (LLM-level risks for multi-turn tool-calling/agents).
- Multi-turn Safety Evaluation (dynamic and multi-turn alignment benchmarks).
- Technical Mastery: In-depth knowledge of machine learning principles and frameworks (PyTorch preferred) with strong Python programming skills.
- Multimodal Systems: Experience working with large multimodal datasets and multi-modal foundational models.
- Soft Skills: Outstanding analytical problem-solving abilities paired with excellent collaboration and communication skills.
- Cultural Alignment: Demonstrates behaviors that build trust: humility, transparency, respect, and intellectual honesty.
Nice to have
- Academic Track Record: Published papers on AI Safety, alignment, or machine learning security as a primary author at top-tier conferences (NeurIPS, ICML, ICLR, ACL, etc.).
- Community Contributions: Active contributions to open-source AI Safety tools, benchmarks, datasets, and/or models.
- Advanced Alignment: Proven experience with alignment/fine-tuning of Vision-Language Models (VLMs) or any-to-text foundational models.