Hello World. My name is
Joseph Ma
And I'm
Résumé

About me

Hello there. I'm Joseph!

I am currently pursuing my master's degree in Electrical and Computer Engineering at Purdue University ECE. While my core focus is embedded systems and firmware, I often work across the stack, whether that means stepping into software development and networking or digging into circuit design and hardware-level debugging.

Most recently, I was at Amazon Robotics as an Embedded Firmware Engineering Co-op working on Proteus, their first autonomous robot. Before that, I interned with the Walt Disney Company as an Attractions Engineer, where I worked on IoT and embedded systems for live deployment on one of their rides. At Purdue, I also serve as a Graduate Teaching Assistant for ECE 362: Microprocessor Systems and Interfacing, where I guide students through labs using ARM-based microcontrollers and embedded C.

As for my personal projects, many of them branch beyond embedded systems into areas like machine learning which is a topic I've been diving deeper into. Want to see more? Check out the Projects section for detailed writeups, interactive demos, and a showcase of 13+ completed projects, with more on the way.

Beyond tech, I find joy in activities such as biking, swimming, watching movies, and occasionally expressing my creativity through digital art + UI/UX design. Feel free to visit the Memories section, where you can discover more about what I've been up to and even set a new homepage background.

Experience

Amazon Robotics

Embedded Firmware Engineering Co-op

January 2025 - July 2025

I joined the Mobile Robotics Safety Firmware (MoRSaf) Team working on Amazon's first autonomous robot, Proteus. My work centered around developing embedded firmware in a real-time operating system (RTOS) environment and managing communication and data transfer between the RTOS and Linux-based systems on a multi-core SoC. This included developing Linux user space applications to initialize cross-core communication, and using JTAG to debug and validate firmware behavior on the RTOS core during bring-up and testing. In addition to firmware development, I built a Python-based gRPC client and CLI to help engineers quickly access diagnostics and health data, cutting on-call query times by 20×. I also added socket fallback for reliability and built in SSH tunneling to support remote querying.

Placeholder Image

Purdue University

MS in Electrical and Computer Engineering

August 2023 - Present

Course Highlights


Project Highlights

  • Cat and Mouse 2 - Reinforcement Learning: Trained an RL agent to outsmart the A* algorithm.
  • Neural Network Architecture: Built and hand derived a neural network from scratch.

Graduate Teaching Assistant

Placeholder Image

The Walt Disney Company

Attractions Engineering Intern - IoT and Embedded Systems

May 2024 - Aug 2024

I led the development and deployment of a real-time monitoring and data acquisition system for a Disneyland boat ride, capturing over 16 critical measurements, including location, speed, water temperature, and fuel levels. I integrated LoRa technology and programmed a gateway optimized for continuous data transmission to a central NAS (Network Attached Storage). This setup enabled 24/7 remote monitoring and prepared the data for Machine Learning to predict and prevent engine failures, potentially reducing downtime by 50% and allowing for an additional 23,000 guests annually. Throughout the project, I refined the design through multiple iterations of breadboard prototypes, ultimately finalizing it on a PCB for mass production. Additionally, I proposed the use of Time Division Multiplexing to enhance transmission efficiency, laying the groundwork for further optimization. By the end of my internship, this design was successfully deployed on 22% of the ride, paving the way for full deployment across the entire attraction.

Placeholder Image

Purdue University

BS in Computer Engineering

August 2020 - May 2023

Course Highlights


Project Highlights

  • Cat and Mouse: An embedded systems game boy featuring an interactive Dijkstra's driven opponent.
  • Audio Equalizer: Designed a circuit to isolate and amplify signals w/ bass, mid, and treble frequencies.

Undergraduate Teaching Assistant

Placeholder Image

Preface

Full Stack Development Intern

June 2021 - August 2021

I redesigned and fixed the web portal interface, making it responsive and cross-browser compatible using HTML5, CSS3, and JavaScript—skills I self-learned during the internship and later used to build this personal website and portfolio. I also optimized backend processes by testing CRUD operations and integrating RESTful APIs with Postman. Additionally, I implemented data validation and security measures within a Ruby on Rails environment using Active Record validations. On the side, I evaluated the machine learning curriculum offered at Preface and provided feedback. This experience sparked my interest that eventually led me to pursue more projects in this field over the years.

Placeholder Image
EMBEDDED SYSTEMS
Cat and Mouse
A game boy programmed in C using the STM32 ARM microcontroller. Utilized SPI, ADC, DAC and GPIO for hardware operation. Cat was programmed to autonomously chase the mouse using Dijkstra's algorithm. Design was finalized with a PCB.
GROUP PROJECT
Cat and Mouse
The Cat and Mouse project is an advanced interactive electronics project featuring a custom-designed PCB integrated with an STM32F091RCT6 ARM microcontroller, a 3.2" ILI9341 TFT LCD touchscreen, and audio output capabilities. Utilizing Dijkstra’s algorithm, the game dynamically calculates the path of an AI-controlled cat seeking the user-directed mouse within a maze. The project employs SPI for efficient LCD communication, along with ADC and DAC for touchscreen interfacing and audio management respectively.
MACHINE LEARNING
Cat and Mouse 2: Reinforcement Learning
This project brings the Cat and Mouse project to another level by training the cat and mouse with reinforcement learning to interact against each other. Initially, the cat uses Dijkstra's algorithm, but both are later trained with Q-learning. Users can design custom maps and train mice against an upgraded cat equipped with A* pathfinding.
INDIVIDUAL PROJECT
Cat and Mouse 2: Reinforcement Learning
This sequel to the Cat and Mouse project uses RL and tabular Q-learning within a Markov Decision Process to train the cat and mouse. Initially, both the cat and mouse are trained against each other, with generational training improving their strategies. A user interface was built to allow users to design custom maps and train mice against an upgraded cat equipped with A* for faster pathfinding. A double depth-first search (DFS) algorithm memoizes dead-end spots, speeding up the mouse's learning and accounting for edge cases. This project has applications in autonomous vehicle navigation, police/criminal car pursuits, navigating dangerous environments (e.g., environmental disasters, wildlife areas), and military encounters.
MACHINE LEARNING
Neural Network Architecture
Built a neural network from scratch by hand-deriving and implementing all calculations, including matrix multiplications, forward and backward propagation, gradient descent, and partial differentials for adjusting weights and biases. The entire algorithm was then written in Python using only NumPy and successfully applied to datasets like MNIST.
INDIVIDUAL PROJECT
Neural Network Architecture
Neural networks often seem like a black box. This project opens the box by constructing one from scratch. Every calculation and algorithm was hand-derived, including matrix multiplications for passing data through layers, forward propagation for neuron activations via weighted sums and activation functions, and deriving gradients of the loss function through backward propagation using the chain rule. Gradient descent iteratively updated the network's weights, minimizing the loss function with hand-derived gradients. Partial differentials were calculated to find the derivatives of the loss function with respect to each weight and bias. The entire neural network was coded from scratch using only NumPy for matrix operations. This hands-on approach successfully trained and tested the network on the MNIST dataset, offering a solid introduction to deep learning.
DATA SCIENCE
Wordle Solver
A Python-based Wordle solver that leverages the Pandas library to efficiently filter and suggest optimal guesses based on color-coded feedback from the game.
INDIVIDUAL PROJECT
Wordle Solver
This Python-based Wordle solver utilizes the Pandas library to create an effective game-solving environment by maintaining a database of potential words and dynamically updating this list based on user input. The solver interprets color-coded feedback to adjust the list of possible words, applying logical conditions: removing words that don't match green and yellow criteria and excluding letters indicated by gray unless they appear as yellow or green elsewhere in the word. It also performs statistical analysis to determine the most informative starting word by calculating the frequency and placement of letters in remaining words after each round. Through this analysis, I identified the best and worst opening guesses for Wordle; the top 25 and worst 25 opening guesses are detailed in my blog post.
ELECTRONICS DESIGN
Goodnight Light
Designed and created a PCB circuit schematic/board layout on Autodesk EAGLE and soldered on components. Final product was a device that turned on/off an LED based on how much light was detected in the environment.
INDIVIDUAL PROJECT
Goodnight Light
The Goodnight Light is an innovative lighting solution designed using Autodesk EAGLE, featuring a circuit that utilizes an LM358 operational amplifier as a comparator to control an LED based on ambient light. The circuit adjusts the LED's state through a high-impedance voltage divider system, where a photoresistor alters its resistance in response to light levels, toggling the LED on in darkness and off in light. This setup ensures energy-efficient operation and offers a practical application in environments where automatic lighting adjustments are beneficial, such as in bedrooms or hallways.
MACHINE LEARNING
K-means Clustering: A Visualization
Wrote an algorithm to illustrate unsupervised machine learning on python in both 1 and 2 dimensions without the help of any machine learning libraries.
INDIVIDUAL PROJECT
K-means Clustering: A Visualization
This project written in Python utilizes the random library to generate coordinates and Matplotlib to visualize the process of K-means clustering. The program plots 120 points on a 100x100 coordinate plane and groups them into six clusters by calculating distances to dynamically repositioned centroids. This visualization serves to demonstrate the K-means algorithm, a machine learning method that optimizes clustering based on minimizing distances within each cluster.
COMPUTER-AIDED DESIGN
Stark Tower
A detailed CAD model of Stark Tower created and designed on CATIA V5. 3D printed with PLA afterwards.
INDIVIDUAL PROJECT
Stark Tower
This CAD model of Stark Tower, inspired by the Avengers movies, is created in CATIA V5 after closely studying images of the iconic building and later printed in PLA. The design is divided into five main components: the roof, circular balcony, Stark display, main body, and bottom base, each showcasing detailed CAD modeling techniques. The model is not to scale, but captures the essence of Stark Tower—download the STL file if you're interested in 3D printing your own version.
ELECTRONICS DESIGN
Audio Equalizer
Designed and constructed an audio equalizer comprising of filters and operational amplifiers to isolate and magnify signals at bass, mid and treble frequencies.
INDIVIDUAL PROJECT
Audio Equalizer
The project focused on building an audio equalizer system capable of fine-tuning bass, mid, and treble frequencies within an input signal. By employing low-pass, high-pass, and band-pass filters, the signal was divided into distinct frequency bands. Operational amplifiers, along with potentiometers serving as adjustable feedback resistors, enabled users to control the scaling and gain of each frequency band. Component values were carefully calculated to meet specific output voltage requirements across various EQ settings and frequencies.
ENVIRONMENTAL ENGINEERING
Weight Amplifier & Comparator
Developed a waste differentiation system using the weight amplifier and comparator to distinguish paper towels from other waste based on their absorbent properties. This system measures the weight gain after soaking waste with water to accurately sort paper towels for recycling, significantly reducing landfill contributions.
GROUP PROJECT
Weight Amplifier & Comparator
Developed a waste differentiation system using the weight amplifier and comparator to sort paper towels from other waste based on their absorbent properties. The system detects, compresses, weighs, soaks, and reweighs waste, calculating the gain (AW) as the weight ratio before and after soaking. High AW values indicate paper towels, which are then sorted for recycling. Initial tests showed high accuracy, highlighting the system's potential to improve recycling efforts and reduce landfill contributions.
EMBEDDED SYSTEMS
Security Alarm Sensor
Codesigned and programmed a security alarm sensor with a PLC. Utilized diffused photoelectric sensor and inductive sensor for detection. Flowchart logic was implemented on nanoNavigator and design was prototyped in SolidWorks.
GROUP PROJECT
Security Alarm Sensor
This security alarm sensor project uses nanoNavigator software and a programmable logic controller (PLC) to implement flow chart logic with simplified block code. It's designed to detect movement and identify potentially dangerous objects, utilizing inputs like a photoelectric sensor, inductive sensor, and push button, with outputs including a blue light, orange light, and buzzer. The outer casing of the device was also designed in SolidWorks
MACHINE LEARNING
Enzyme Data Analysis
Developed a model to accurately fit data of enzyme reactions, analyzed reaction velocities and substrate concentrations using the Michaelis-Menten model to find the best-fitting parameters, Vmax and Km, for five enzymes. The project involved nonlinear regression, combining algebraic methods and gradient descent to optimize accuracy.
GROUP PROJECT
Enzyme Data Analysis
This project determined optimal parameters, Vmax and Km, for enzyme kinetics using the Michaelis-Menten equation. We analyzed five enzymes with duplicate tests, generating 10 data sets. Initial reaction velocities were calculated from substrate concentrations by determining slopes on the [P] vs. Time plot. Nonlinear regression was applied, first using algebraic methods to estimate Vmax and Km, then testing all possible parameter combinations (2025 per enzyme) to minimize Sum of Squared Errors (SSE). Gradient descent refined these parameters by exploring ±10% variations, ensuring convergence to global minima. This comprehensive methodology accurately modeled enzyme kinetics.
CNC MACHINING
CNC Machined Hammer
Generated G-code using Vericut for CNC machining and performed the final operation on a Hurco CNC machine. The project involved engraving over an area of 2.5 * 0.6 inches on a hammer head and included various commands for different machining operations, such as linear patterns and round corners.
INDIVIDUAL PROJECT
CNC Machined Hammer
This project involved generating G-code for CNC machining using Vericut and performing the final operation on a Hurco CNC machine to engrave a hammer head over an area of 2.5 * 0.6 inches. The initial G-code was written in a text file and then uploaded to Vericut for simulation. Commands included setting the reference zero origin, loading tools, starting the spindle at specific RPMs, and executing rapid traverse motions. Linear motions and radial arcs were used to create patterns and engravings with precise feed rates. Toolpaths were designed for linear patterns and round corners, requiring careful adjustments to ensure accurate engravings. The final design featured the Avengers logo and my initials on each side of the hammer head.
WEB DEVELOPMENT
Multivariate Grade Calculator
My first web dev project. The Multivariate Grade Calculator is a web-based application built with HTML, CSS, and JavaScript, designed to dynamically calculate and visualize academic performance across various assessments.
INDIVIDUAL PROJECT
Multivariate Grade Calculator
The Multivariate Grade Calculator began as a Python-based tool using Matplotlib and object-oriented programming to manage and analyze grades and evolved into a web-based application built with HTML, CSS, and JavaScript. It offers a dynamic user interface that updates in real time to reflect changes in assessment data, maintaining user data across sessions via local storage. Interactive elements like sliders allow users to adjust potential scores, and a unique Grade Thermometer visually represents the maximum and minimum achievable grades based on current inputs. This enables students to strategically prioritize certain assessments over others, depending on their impact on the final grade, thus optimizing their academic efforts for the best possible outcomes. (Responsive version is still in development).
ROBOTICS
Gyroscope Controlled Arm
Built a robotic arm out of popsicle sticks/MG966R servo motors and used Arduino software to calibrate them to GY-521 gyroscope accelerometers which allowed for the robotic arm to mimic the movements of a human arm.
INDIVIDUAL PROJECT
Gyroscope Controlled Arm
This project involves constructing a robotic arm from popsicle sticks, controlled by an Arduino Mega 2560 and programmed using the Arduino Software (IDE). The arm's movements are finely tuned through a GY-521 MPU6050 module, which integrates an accelerometer and gyroscope to accurately map the servo motors' positions. A Hall Effect Sensor is also employed to enable the gripper to respond to magnetic fields, enhancing the arm’s interactive capabilities.
Copyright © 2022 - 2025 Designed & Built by Joseph Ma