Decision Intelligence Engineer - Action

Position: Decision Intelligence Engineer - Next Best Action
** Become a part of our caring community*
* Become a part of our caring community and help us put health first. We are looking for a skilled Decision Intelligence Engineer to design, train, and improve the reinforcement learning policy at the heart of Humana's Next Best Action platform.

This role is hands-on and research-oriented. You will design and evaluate decision-making algorithms, and instrument training pipelines. Additionally, you will collaborate with data and platform engineers. Furthermore, you will ensure the system operates correctly within the constraints of clinical eligibility rules and program-specific objectives.

** KEY RESPONSIBILITIES*
* ** Decision-Making Model Development*
* + Design, implement, and evaluate algorithms suited to long-horizon, sparse-reward sequential decision-making in healthcare. These algorithms include reinforcement learning methods, such as PPO, A3C, DQN, CQL, and Decision Transformer, as well as dynamic programming formulations and constrained optimization approaches.

+ Frame member decisioning problems as Markov Decision Processes (MDPs) or Partially Observable MDPs, defining state representations, action spaces, transition dynamics, and reward structures that encode clinical and program-specific goals.

+ Apply Bellman-equation-based value estimation, reward shaping, and constraint formulations to encode clinical eligibility, suppression rules, and program-specific objectives directly into the learning or optimization objective.

+ Manage exploration-exploitation tradeoffs (or equivalent uncertainty-handling in simulation and stochastic optimization) appropriate for a production healthcare environment where suboptimal actions have member impact.

+ Model member journey dynamics using tools from stochastic processes, simulation, or probabilistic graphical models to inform policy design and evaluate.

** Model Evaluation and Production Safety*
* + Build simulation and backtesting environments, including discrete-event simulation and Monte Carlo methods, to evaluate policy or decision quality before production promotion using historical member journey data.

+ Diagnose failure modes specific to learned or optimized policies. These include policy collapse, credit assignment errors across long member journeys, distributional shift between training and serving populations, and constraint violations under out-of-distribution inputs. Remediate these failure modes.

+ Define performance threshold criteria and automated evaluation gates within the nightly Databricks training workflow; block promotion of underperforming policies to MLflow production.

+ Instrument training and optimization runs with MLflow tracking covering hyperparameters, objective curves, action distributions, and feature importance for every training cycle.

** Training Pipeline Engineering*
* + Own the nightly Databricks training workflow. This workflow involves feature engineering from upstream clinical and operational data sources, and state vector normalization. Additionally, it includes distributed training by Ray RLlib (or equivalent optimization solvers), and batch scoring of all eligible members.

+ Collaborate with the Data Engineering team to ensure the Data Engineering team correctly joins training inputs, computes reward signals from disposition outcomes, and makes the feature pipeline reproducible and auditable.

+ Write production-quality PySpark feature engineering jobs; maintain data lineage through Databricks Unity Catalog.

+ Manage model artifacts, versioning, and lifecycle in the MLflow Model Registry; ensure rollback capability is maintained at all times.

** Multi-Agent and Constraint-Aware Decisioning*
* + Apply multi-agent decision-making concepts (MARL via Petting Zoo, or game-theoretic or cooperative optimization approaches) where member household or population-level coordination is required.

+ Implement constraint handling to enforce hard business rules directly within the optimization objective. These rules include member caps, cooldown periods, and clinical eligibility. To achieve this, use constrained MDP formulations, Lagrangian relaxation, or mixed-integer programming as appropriate, rather than relying on downstream filters.

+ Collaborate with rules engine stakeholders to ensure eligibility guards and policy priorities are correctly aligned and do not conflict.

** Collaboration and Governance*
* + Partner with decision engine and rules engine teams to ensure that you integrate model outputs cleanly with the real-time decisioning hot path and that you correctly structure and interpret scored recommendations.

+ Collaborate with platform architects to define feedback loop contracts: how disposition outcomes flow back through the data pipeline into the next training cycle.

+ Document model behavior, known limitations, and failure modes for clinical and compliance stakeholders; support explainability requirements for member-facing decisions.

+ Use AI-assisted engineering tools for scaffolding, testing, and…