Reinforcement learning engineer: agents for control and optimization

I develop reinforcement learning solutions for control, optimization and decision-making tasks and lead a project from a vague idea to an agent that stably improves your process. Work almost always starts with the toughest and most underestimated phase — correct problem formulation. Together we define state, actions and, crucially, reward because a poorly designed reward function produces an agent that games the metric instead of solving the task. I design environments and simulators close enough to your real dynamics to train in them and verify that behavior learned in simulation actually transfers to production.

With algorithms I start from your task, not from fashion. For continuous control I usually pick PPO or SAC, for discrete and combinatorial solutions — DQN and modern variants, and in contextual settings carefully tuned bandits are more honest than overcomplicated architecture. I close the full training cycle: vectorized environments, reward shaping, curriculum, hyperparameter search and honest evaluation against strong baselines so you see real gains, not a hand-picked pretty curve.

I treat deployment separately because that's where most RL projects quietly die. I export the policy in a light format for inference, add safety constraints and fallback scenarios for unusual states, set up monitoring and retraining as data drifts. You get a working agent, clean code, documentation and project handoff to your team.

— Environment design and reward functions, simulators
— PPO, SAC, DQN, contextual bandits, offline RL
— Training, tuning, comparison with baselines
— Deployment, safety constraints, monitoring