Inspiration
This past year at Brown University, I had the opportunity to serve as a volunteer transporter at the Miriam Hospital (a local Providence hospital), and was responsible for delivering patients, samples, and belongings to their proper locations. This experience was extremely eye-opening for me, as it gave me my first real understanding of how hospitals operated, and also revealed inefficiencies in routine clinical workflows. Particularly, I noticed how much time was spent on unnecessary manual checks of paperwork and specimen bins in hospital wards, often just to find that nothing had been dropped off or picked up. This would take away valuable time from other tasks such as discharging patients or transporting belongings. Other times, samples and paperwork would sit in bins for extended periods and risk being delayed, which could hinder downstream processing. While I observed these issues firsthand at Miriam Hospital, I realized that the underlying challenge was broader: many healthcare environments still rely on repetitive manual status checks and fragmented communication processes that can create inefficiencies in day-to-day operations. These observations ultimately inspired Warden, a system designed to replace repetitive status checking with real-time visibility into hospital workflows.
What it does
The current prototype is a smart bin designed to serve as a central transport hub within a hospital ward, reducing the need for repetitive manual status checks. When materials such as specimens or paperwork are placed into the bin, a camera compares the current status with a baseline image to detect whether or not the bin is occupied. Additionally, using an AI-powered computer vision model trained on hospital workflow materials, Warden distinguishes whether the item is a specimen, paperwork, or multiple items simultaneously and updates a centralized dashboard in real time. The dashboard provides staff with immediate visibility into bin status, item contents, and how long materials have remained in the bin. By minimizing repetitive tasks and improving workflow visibility, Warden helps return valuable time and attention to where it matters most: patient care.
How I built it
First, I built the physical prototype itself. For the current version, the smart bin measures approximately 5 × 10 × 12 inches and is constructed using wooden boards. The interior is fitted with an LED light strip to provide consistent lighting conditions and improve detection accuracy, while a USB camera is mounted in the corner to continuously monitor the bin contents. Altogether, the materials required for the prototype totaled under $50, making Warden a relatively low-cost and scalable approach for improving hospital workflow efficiency. For the software component, I initially focused on solving a simpler problem: determining whether the bin was occupied or empty. To accomplish this, I developed a computer vision system that compared the live camera feed against a baseline image and measured changes within the bin environment. After establishing reliable occupancy detection, I expanded the system by creating and labeling several hundred images in a custom dataset using Roboflow. These images were used to train an AI-powered object detection model capable of distinguishing between hospital workflow materials such as specimens and paperwork. I then integrated the trained model into a real-time monitoring system that continuously analyzes incoming camera data, identifies the contents of the bin, and updates a centralized dashboard displaying occupancy status, detected materials, and item duration within the bin. Building this prototype required combining physical prototyping, computer vision, machine learning, and software development into a single integrated system designed around a practical challenge observed in healthcare environments.
Challenges I ran into
The first challenge involved unstable detection under inconsistent or low-light conditions. Changes in ambient lighting significantly affected the camera feed and reduced the reliability of image analysis. To address this, an LED light strip was installed inside the bin to create a controlled and consistent lighting environment, improving the stability and accuracy of detection. Another challenge was that even very small shifts in camera position, environmental changes, or random image noise could trigger false positives and incorrectly classify the bin as occupied. Initially, the system was highly sensitive to small pixel-level changes such as shadows, lighting fluctuations, and camera noise. To address this, I implemented a baseline comparison system that continuously compares the live camera feed against a reference image of an empty bin. Gaussian blur was applied before image comparison to smooth the frames and reduce sensitivity to minor visual noise, while a smoothing mechanism averaging changes across multiple frames prevented single-frame spikes from triggering detections. Additionally, the baseline image automatically updates after the bin remains empty for a period of time, allowing the system to adapt to gradual environmental or camera changes and maintain long-term reliability. A further challenge involved unstable AI predictions. Early versions of the system attempted to classify objects immediately after they entered the bin. However, items were often still moving or had not yet settled into their final position, causing the model to occasionally misclassify objects, rapidly switch between labels, or become stuck in an "unidentified" state even after the contents later became recognizable. To improve reliability, I implemented a short cooldown period before running AI detection, allowing materials time to settle before analysis occurred. Confidence thresholds were also introduced so that predictions would be based on stronger detections rather than uncertain frames, and the detection pipeline was modified to continuously re-evaluate bin contents at regular intervals rather than assuming the first prediction was final. This allowed the system to recover from uncertain detections and dynamically update as materials entered, moved within, or left the bin.
Accomplishments that I’m proud of
Aside from developing solutions to the challenges encountered throughout the project, I am most proud of taking an inefficiency I observed in a real healthcare environment and transforming it into a functioning prototype with practical applications. Rather than creating a purely theoretical solution, Warden was inspired by firsthand observations from hospital workflows and built around a specific challenge affecting day-to-day operations. I am also proud of successfully integrating multiple disciplines into a single system. Warden combines physical prototyping, computer vision, machine learning, real-time monitoring, and dashboard development into one cohesive workflow. Building a system in which the hardware and software components communicate reliably required extensive iteration and problem solving throughout development. Finally, I am proud of how Warden evolved throughout development. What began as a simple occupancy detector ultimately developed into an adaptive system capable of identifying materials, tracking status in real time, and providing a more intelligent approach to workflow monitoring.
What I learned
I learned how much trial and error can go into addressing problems that initially seem simple or straightforward. Warden was my first experience designing a system that combined both physical hardware and software into a single functioning product, requiring me to think not just about individual components, but about how they interact as an integrated system. What initially seemed like a straightforward idea quickly expanded into solving challenges involving lighting conditions, camera positioning, AI reliability, and overall system stability. This experience showed me that effective solutions are built through continual testing, iteration, and refinement.
What's next for Warden
One of the next priorities for Warden is improving detection performance when multiple objects are present simultaneously within the same bin. While the current system can identify specimens and paperwork, performance becomes less reliable when both are present together, occasionally causing the model to classify the entire scene as a single object type. Expanding and diversifying the training dataset with additional images containing multiple materials, varying object positions, partial overlap, and different lighting conditions would help improve the model’s ability to distinguish multiple items more consistently. Improving this capability would also move Warden closer to functioning as a true central transport hub within a hospital ward. Currently at Miriam Hospital, paperwork and specimens are collected at two different locations within each ward, requiring volunteers to monitor multiple collection points. A system capable of reliably tracking multiple materials within a shared space could help consolidate these workflows and provide a more centralized view of transport activity. Beyond technical improvements, another major next step is exploring real-world implementation. Since Warden was inspired by firsthand experiences within a hospital environment, I hope to discuss the concept with hospital staff and volunteer leadership to better understand practical considerations and gather feedback on how the system could be integrated into existing workflows. While initial conversations would likely begin within the Miriam Hospital, I would aim to expand testing and feedback to additional hospitals and healthcare settings in order to evaluate how Warden performs across different environments. Moving from a prototype environment into real clinical settings would provide valuable insight into how the system could continue evolving to address broader operational challenges in healthcare.
Log in or sign up for Devpost to join the conversation.