This paper presents an adaptable, safe, and portable robot-assisted feeding system designed to help people with mobility impairments feed themselves. The system is designed with three key principles: safety, portability, and user control. It includes comprehensive safety checks, the ability to be mounted on and powered by any powered wheelchair, and a custom web app that allows users to control the robot using their own assistive devices. The system has been validated through multiple end-user studies and has successfully assisted individuals with severe medical conditions such as Multiple Sclerosis, Spinal Cord Injury, Spinal Muscular Atrophy, and Arthrogryposis. The system uses two robot arms, the Kinova Gen2 (JACO) and the Kinova Gen3, both with 6 degrees of freedom. The JACO has an eye-in-hand system that allows for continuous wrist joint rotation, while the Gen3 has a second camera. The system uses a custom 3D-printed fork assembly with a 6-DOF ATI Nano25 force-torque transducer. The primary compute is a Lenovo Legion 5 laptop with an Nvidia RTX 3060 6GB GPU, and the primary networking component is a Cradlepoint IBR900 router, both mounted on the back of the user's wheelchair. The software stack is built on ROS2 and the ros2-control framework. The system uses a contextual bandit framework augmented with haptic post hoc context to achieve online adaptability. The robot will try different discrete actions with new food items and will figure out the best one over the course of 8-13 attempts. For bite transfer, the robot uses real-time mouth perception and interaction-aware control to deliver food to the care-recipient's mouth. The system has been tested through end-user studies and deployments over several years, demonstrating its effectiveness in real-world scenarios.This paper presents an adaptable, safe, and portable robot-assisted feeding system designed to help people with mobility impairments feed themselves. The system is designed with three key principles: safety, portability, and user control. It includes comprehensive safety checks, the ability to be mounted on and powered by any powered wheelchair, and a custom web app that allows users to control the robot using their own assistive devices. The system has been validated through multiple end-user studies and has successfully assisted individuals with severe medical conditions such as Multiple Sclerosis, Spinal Cord Injury, Spinal Muscular Atrophy, and Arthrogryposis. The system uses two robot arms, the Kinova Gen2 (JACO) and the Kinova Gen3, both with 6 degrees of freedom. The JACO has an eye-in-hand system that allows for continuous wrist joint rotation, while the Gen3 has a second camera. The system uses a custom 3D-printed fork assembly with a 6-DOF ATI Nano25 force-torque transducer. The primary compute is a Lenovo Legion 5 laptop with an Nvidia RTX 3060 6GB GPU, and the primary networking component is a Cradlepoint IBR900 router, both mounted on the back of the user's wheelchair. The software stack is built on ROS2 and the ros2-control framework. The system uses a contextual bandit framework augmented with haptic post hoc context to achieve online adaptability. The robot will try different discrete actions with new food items and will figure out the best one over the course of 8-13 attempts. For bite transfer, the robot uses real-time mouth perception and interaction-aware control to deliver food to the care-recipient's mouth. The system has been tested through end-user studies and deployments over several years, demonstrating its effectiveness in real-world scenarios.