Adam Del Rosso

Software Engineer

AWS Solutions Architect

Neurotech Researcher

Resume Contact Me
About Me

Hi! My name is Adam Del Rosso.

I'm a Full Stack/IoT Software Engineer at Capstone Integrated Solutions and recently graduated from the Rochester Institute of Technology with a degree in Software Engineering. For the past few years, I've been developing a specialty in cloud-native systems architecture and development. The power and flexibility of cloud-native systems continue to impress me, and I enjoy using the cloud toolbox to build new systems or give new life to existing ones.


Outside of work you'll find me contributing to the open source projects we all depend on, biking, hiking, golfing, and otherwise trying to stay active in life away from the screen.

Notable Achievements
RIT Business Model Competition 2020 - 1st
Project featured on RIT magazine cover
Computer Science II Game AI - 2nd of >100
Eagle Scout
Certifications

AWS Certified Solutions Architect - Associate AWS Certified Developer - Associate AWS Certified Cloud Practitioner
Recent Works
Notes App
AWS

Notes App

AWS

I created a simple note-keeping web app called “scratch” using a stack of serverless technology.

The heart of the backend is a serverless REST API that utilizes AWS API Gateway and AWS Lambda. Lambda “functions” are short-living compute containers that are started in a response to some trigger, in our case an API call. The application also relies on Amazon Cognito for user management and authentication, Amazon Simple Storage Service (S3) for user file uploads, and Amazon DynamoDB for user note data. These services are managed with configuration files and use the Serverless application framework to manage the provisioning of services.

The front end is a single page application (SPA) written in JavaScript using the React framework. SPAs remove the need for page loads by delivering the code for the entire application upon first navigating to the website. This enables more responsive web applications that can rival native apps. The frontend data is hosted on Netlify’s content distribution network (CDN) which removes the concern of managing a dedicated web server.

Thanks to these serverless technologies, I do not need to worry about the availability or scalability of my application, as it is in the hands of the service providers. Although the project has little traffic, the backend is capable of smoothly scaling up to meet the needs of millions of users without delays or slowdowns, and without requiring any infrastructure management. As for cost, the application could support hundreds of users within the limits of the AWS free tier.

Future plans

I hope to further improve the user experience of this app by utilizing service workers, which will allow note data to be cached on user devices and displayed immediately once opening the app. This improvement will help bring the web app closer to a native app experience.

Furthermore, I intend to make the service end-to-end encrypted, meaning that user data is encrypted by the client app before it is stored within the cloud services. This way the developer (me) is completely unable to view the notes or files stored by the user. The current version uses TLS to protect the user data in-transit, along with encryption at rest for S3 and DynamoDB.

This project was created following the in-depth Serverless Stack guide.

View Project
Thought-Controlled Wheelchair
Neurotech

Thought-Controlled Wheelchair

Neurotech

This project was featured on the cover of the Spring 2020 RIT University Magazine.

I am leading a research team that is working to develop a brain-computer interface (BCI) for the control of a wheelchair. While BCIs have applications in many domains (cognitive therapy, gaming, sleep, and meditation, among others), our focus is on restoring independent mobility to individuals whose needs are not met by traditional control mechanisms. This system is particularly useful for individuals with locked-in syndrome who have no control over voluntary muscles and may have limited eye movement. This project runs within the Neurotechnology Exploration Team (NXT) at RIT, which is the college’s first undergraduate research group.

The BCI is based on a non-invasive brain activity recording technique called electroencephalography (EEG). EEG utilizes electrodes placed on the scalp to record the weak electrical signals produced by large populations of neurons within the outermost layers of the brain (measured in microvolts). With training, users can learn to modulate the activity of certain regions of the brain through forms of imagery; our group is primarily focused on motor imagery (imagined movement of limbs), however, we are also exploring the use of cognitive tasks like mental arithmetic. Over time, many users are able to modulate their brain activity to a sufficient degree so that signal processing and machine learning algorithms may distinguish between a few imagery tasks. At the same time, machine learning algorithms are continually adjusting to improve the accuracy of the BCI, creating what is called a coadaptive BCI system.

EEG-BCI systems are difficult to engineer in part due to the high levels of noise in the signal. Sources of noise include the AC electrical systems in the room, electrical activity from muscles, movement of the eyes, blinking, and movement of electrodes; electrode movement is made worse by the fact that a wheelchair is a moving platform that might not be traveling on a smooth surface. Due to these challenges and others, even the best EEG-BCI systems used in controlled environments struggle to reach accuracy in the 90% range, and can vary significantly between users. Combine this less than perfect accuracy with the common reports of mental fatigue in users, and the BCI begins to look inadequate for the task of direct wheelchair control, especially considering that the safety of the user and those around them may be put at risk. These limitations lead to the conclusion that an intelligent semi-autonomous wheelchair system must be created before an EEG BCI is suitable for long term use in an uncontrolled environment.

Because the development of this semi-autonomous system falls out of the scope of our research group, I connected with Sidney Davis, president of the RIT Multidisciplinary Robotics Club (MDRC). Sid decided to take on the project, and made a proposal to work on it within the Kate Gleason College of Engineering’s Multidisciplinary Senior Design program. I now serve as the customer for this senior project, and Sid is leading a group of four other seniors on its development. It is our goal to be testing functionality on an integrated BCI and wheelchair system in the Spring of 2021.

See more about NXT at nxtr.org

Zoo Stream Animal Detector
AWS

Zoo Stream Animal Detector

AWS

Header photo credit to the San Diego Zoo.

As part of the RIT class Engineering Cloud Software Systems, teams were asked to propose and develop a project that showcases the capabilities of the AWS Cloud. Our team chose to develop a system that notifies a set of subscribers when an animal becomes visible in a Zoo livestream. With our system, rather than occasionally checking live streams and hoping to see something interesting, subscribers will know exactly when their favorite animals are out and about. Although not limited to this specific exhibit, our proof-of-concept was developed using the polar bear exhibit from the San Diego Zoo.

Architecture

I led the design and development of the system for our team. The system is primarily enabled by the Rekognition CustomLabels service, which is a managed machine learning image recognition offering. CustomLabels accepts a dataset of labeled images from which it trains a machine learning model to (hopefully) a high degree of accuracy. The trained model may then be deployed into a managed inference service capable of labeling multiple images per second. We use SageMaker Ground Truth to develop a set of labeled training images using the provided web application and private worker groups. The “workers” in this group log in to the Ground Truth application and label images of the exhibit as either containing a polar bear or not. In a real deployment of this system, we imagine Zookeepers or other Zoo personnel labeling images from live streams on an occasional basis. Finally, our team used Python code in a docker container deployed on Fargate with ECS to read from the stream, call Rekognition to label stream frames, and send notifications to subscribers when appropriate. SNS was used to send subscriber notifications in this proof-of-concept system, however, a real deployment would use a managed email service like Amazon SES.

Zoo Stream Animal Detector Architecture Diagram

View larger image

System Demo