Senior Machine Learning Engineer, Perception and Controls

Capstan Medical, Inc.•West Menlo Park, CA

1d•Hybrid

About The Position

Teach a heart valve robot to see where it is. We're hiring a Senior Machine Learning Engineer to join our robotics and controls team. You'll help our robot understand where its devices are inside the patient, using intraprocedural imaging such as fluoroscopy and echocardiography. That means building perception for image-to-robot registration and pose estimation, and taking it from research into a clinical system. This is not a model-in-a-notebook role. Your work runs on a real-time robotic system, informs what physicians see during the procedure, and ships under FDA design controls. You'll also help shape how we collect, organize, and learn from our robot and imaging data, and use it to improve our device models and controllers. You'll have a large say in how ML is done here, from the tools and standards we adopt to which problems we tackle first, and you'll be the person the rest of the team turns to for ML judgment.

Requirements

MS in Computer Science, Electrical Engineering, Robotics, or a related field. PhD is a plus.
4+ years of professional experience in computer vision or machine learning, classical or learned, including at least one system shipped to production on real-world data.
Experience working with scarce, noisy, or expensive data, using synthetic data, augmentation, self-supervision, or physics priors to make up for it.
Experience building ML on data from physical systems such as robots, cameras, medical imaging, or sensors.
Ability to thrive in a fast-paced environment and own problems end-to-end.
Ability to work onsite 3–4 days per week, split flexibly between our Santa Cruz, CA and West Menlo Park, CA offices, with at least 1 day per week in Santa Cruz.

Nice To Haves

Ultrasound or echocardiography experience, strongly preferred. Fluoroscopy or X-ray experience is also valuable, including 2D/3D registration or digitally reconstructed radiographs.
Experience tracking catheters, needles, surgical tools, or continuum robots in images.
Sensor fusion and state estimation, such as Kalman filters or factor graphs.
System identification or learned dynamics for control, including residual models on top of physics.
Robotics software experience with ROS/ROS2 and on-device inference under latency constraints (ONNX, TensorRT).
Data platform and MLOps experience, including practices like dataset versioning and experiment tracking.
Medical device or safety-critical background, including validating ML under FDA design controls.
Experience as an early ML engineer on a team, setting up the tools, data practices, and evaluation that others build on.
Publications at venues such as MICCAI, IPCAI, CVPR, ICRA, or IROS.

Responsibilities

Develop methods to register intraprocedural imaging, such as fluoroscopy and echocardiography, to the robot.
Estimate the pose of catheters and implants from medical images.
Combine image-based estimates with robot kinematics and other sensing, and quantify the uncertainty of each estimate.
Build data pipelines that capture, align, label, and version imaging and robot data for training and evaluation.
Use simulation and synthetic data where real clinical data is limited.
Fit and improve models of device behavior from data, and work with controls engineers to turn them into better control.
Work with clinical engineers and physicians to learn what the images show, what matters in the procedure, and what accuracy is good enough.
Read the literature, prototype what looks promising, and tell us plainly what works on our data and what doesn't.
Deploy models on the robot and meet real-time latency budgets.
Write the validation, test datasets, and design documentation that let learned components pass FDA design controls.