Pathfinder: Repath

Back to Portfolio

Problem

Operating rooms generate massive single-use medical device waste, but no scalable system exists to sort and track high-value devices at disposal. OR nurses are already overburdened, making manual sorting infeasible. Working with ZHAW business students and sponsored by Johnson & Johnson, our Stanford team designed an automated sorting and tracking solution deployable at the point of disposal.

Outcome

Pathfinder is a smart sorting bin that uses dual-camera computer vision to identify single-use medical devices and a motorized paddle-wheel to sort them into high-value and low-value bins in under three seconds. The computer vision model was trained on EdgeImpulse against a standardized background. Pathfinder feeds into the broader RePath ecosystem — a shared platform connecting hospitals, logistics orchestrators, and manufacturers for end-to-end device traceability.

Design

Settling on a solution took months, with multiple iterations even after committing to a sorting bin. Sketches focused heavily on the paddle wheel — our most complex element. The final Pathfinder uses an aluminum extrusion frame with sheet walls, caster wheels, and a custom riveted paddle-wheel driven by a belt transmission. Two 4K USB cameras capture device orientation, and an Arduino system coordinates the motor, cameras, LEDs, RFID reader, and database.
After settling on the smart bin concept, we built an initial CAD model featuring a bag dispenser, camera, heating element, and actuator. This was presented to J&J and ZHAW partners in Switzerland, where stakeholder feedback drove a pivot toward broader supply chain traceability and reduced plastic use.
Engineering sketches compared bevel gear vs. belt transmission options for the paddle wheel. The belt transmission was chosen for its motor placement flexibility and compatibility with the Pathfinder's constrained housing.
Two paddle-wheel configurations were explored in CAD — angled and upright. A quick functional prototype revealed the stapler caught on the bin rim. Feedback from Swiss partners also flagged inconsistent bin sizing across ORs, leading to a lengthened frame for better downstream integration.
I was responsible for the final CAD.

Build

The prototype uses aluminum extrusions and sheet panels, with a vacuum-formed plastic lid for consistent device presentation to the cameras. A belt-driven DC motor powers the paddle wheel, with limit switches defining each sorting cycle. The computer vision model was trained on ~4,000 images of three J&J devices, using bounding boxes focused on distinctive plastic ends for reliable identification. The full-stack software runs on XAMPP with a MySQL database, PHP backend, and HTML/CSS/JS dashboards tailored to each stakeholder.

Testing and Iterations

A cardboard training jig was built early to develop the computer vision model in parallel with hardware. Camera placement, height, and lighting were iterated through hands-on testing. The model was refined across multiple rounds — scaling from ~250 to 1,000 images per device — and reached 89.3% identification accuracy with false positives occurring roughly 1 in 20 attempts. Paddle-wheel geometry was also iterated, including a rejected angled-axis concept for the 22-inch surgical stapler.

Reflections

The team repeatedly had to unlearn assumptions — pivoting first from device sterilization to end-of-life sorting, then from a standalone bin to a full ecosystem after the Switzerland trip. The most impactful decisions came from OR nurses, who made clear they couldn't absorb extra workflow steps. Building hardware, software, and a business model simultaneously under an academic timeline revealed how fast integration complexity compounds.