Experience:

• 4 - 7 Years

Responsibilities:

• Own and drive full RLHF pipeline including data collection, reward model training, and RL fine-tuning
• Design and run Supervised Fine-Tuning pipelines on open-weight models
• Build and train reward models capturing human preferences
• Design human feedback collection pipelines including labeling rubrics and annotator calibration
• Implement Constitutional AI and RLAIF techniques to reduce human annotation reliance
• Red team models post-training to identify jailbreaks, regressions, unsafe outputs, and alignment failures
• Design and maintain evaluation benchmarks for alignment, safety, and capability
• Optimize inference pipelines and runtimes to serve aligned models efficiently at scale
• Implement quantization strategies to deploy fine-tuned models on target hardware
• Write and tune low-level C/C++ and Rust code for inference performance
• Diagnose and resolve training instabilities, reward hacking, and production inference bugs
• Stay updated with latest alignment and RL research and translate findings into experiments

Nice to have:

• Experience with Direct Preference Optimization (DPO) and its variants
• Understanding of reward hacking and Goodhart’s Law mitigation strategies
• Experience diagnosing training instabilities such as reward collapse and mode collapse
• Strong mathematical foundation in RL theory, probability, linear algebra, and optimization
• Familiarity with transformer attention and tokenization
• Experience with distributed training for large-scale fine-tuning
• Experience with vector databases like FAISS or Milvus
• Familiarity with retrieval-augmented generation (RAG) pipelines
• Experience integrating LLMs with external tools, APIs, and agent-based systems
• Exposure to Rapid Application Development (RAD) approaches

More skills:

Reinforcement Learning from Human Feedback (RLHF), Policy gradient methods, Supervised Fine-Tuning (SFT), Reward model training, Human feedback collection, Evaluation benchmarks, Inference optimization, llama.cpp, TensorRT, Quantization (INT4, INT8, FP8, LoRA, QLoRA), C, C++, Rust, Python, Distributed training frameworks, Vector databases (FAISS, Milvus), Retrieval-augmented generation (RAG), LLM integration with external tools and APIs, Rapid Application Development (RAD)

Other:

• Probationary employment