Role: Reinforcement Learning & Post-Training Engineer Role Overview As a founding post-training specialist, you will own the end-to-end alignment and reasoning stack taking base pre-trained models through advanced alignment. You will design and execute novel post-training recipes spanning RLAIF, Constitutional AI, and dense process supervision (PRMs) to unlock frontier reasoning and domain alignment directly on raw base weights. Key Responsibilities • Base Model Post-Training: Design and scale full-spectrum post-training workflows on base foundation models using algorithms like GRPO, PPO, DPO, and iterative rejection sampling. • RLAIF & Constitutional AI: Architect automated preference-data synthesis, iterative multi-turn critique loops, and rubric-driven constitutional refinement pipelines to eliminate human annotation bottlenecks. • Process Reward Modeling (PRMs): Construct and train step-level PRMs to evaluate granular reasoning steps, intermediate chain-of-thought derivations, and dense domain verification tasks. • Verifiable Reward & Environment Design: Implement deterministic verifiers, execution sandboxes for code and math, and hybrid reward functions combining outcome checks with dense step feedback. • Alignment Infrastructure: Optimize distributed RL training and rollout loops across accelerator clusters, maximizing rollout throughput, memory management, and inference-engine integration. Candidate Requirements • 3+ Years Hands-On RL / Fine-Tuning Experience: Demonstrated track record executing post-training on LLM base models (not generic software engineering or high-level API wrapper development). • Algorithmic Depth: Deep theoretical and practical command of modern policy gradient methods, credit assignment, value function estimation, and reward-shaping dynamics. • Dense Reward & Alignment Mastery: Concrete experience designing reward models (ORMs/PRMs), preference optimization objectives, or constitutional self-critique mechanisms. • High-Throughput RL Stack: Strong fluency with distributed training and rollout architectures (such as Ray, vLLM, DeepSpeed, or custom JAX/PyTorch accelerator pipelines). • Startup Velocity & Ownership: Ability to independently formulate research hypotheses, diagnose reward hacking or mode collapse, and deploy robust training runs under fast production cycles. What We Offer • Direct Architecture Ownership: Full technical authority over the post-training, alignment, and reasoning methodology of foundation models. • High-Density Compute: Dedicated accelerator clusters to validate novel RL formulations and train production models end-to-end.