Rehabilitation AI Partner

AI Systems Engineer

A production-oriented JITAI system for adaptive rehabilitation support across long recovery timelines.

The architecture combines longitudinal reinforcement learning with psychological-state detection to personalize intervention timing and intensity rather than relying on static reminders.

Design emphasis is placed on policy behavior over time, balancing adherence support with intervention burden in a system that can be trained safely before deployment.

Deep Reinforcement Learning PPO Transformer NLP Digital Twin JITAI Behavioral Modeling
Rehabilitation AI Partner overview

Problem

Rehabilitation success depends on sustained daily adherence, but most intervention systems use fixed prompting strategies that ignore behavioral drift, psychological state, and delayed consequences of intervention decisions.

Solution

The system is designed as a Just-in-Time Adaptive Intervention engine powered by deep reinforcement learning. A transformer-based NLP classifier detects psychological state from user dialogue, and the output is fused with telemetry to build the decision state for policy selection.

A PPO policy agent then selects intervention intensity from a structured action space: Silence, Nudge, Micro-Intervention, or Active Dialogue. The decision process is framed as a full Markov Decision Process to capture delayed outcomes, with a synthetic User Simulator used for offline policy training prior to real deployment.

System Architecture

Rehabilitation AI Partner system architecture diagram

User Dialogue + Telemetry -> NLP Classifier (Psychological State Detection) -> State Vector Construction -> PPO Policy Network -> Action Selection (Silence / Nudge / Micro-Intervention / Dialogue) -> Intervention Delivery -> Reward Signal -> Policy Update (Offline in Simulator).

Engineering Decisions

Validation & Iteration Strategy

Validation is executed offline in the synthetic environment through scenario-driven policy replay and controlled baseline comparison across random, heuristic, and PPO agents.

Design Constraints

My Role

View Code Back to Projects GitHub Demo