Final year · Stevens Institute of Technology · B.E. May 2027

Lily
Stone

Computer engineer, one year out. I founded Amplifly and put it on the App Store, designed the recommendation algorithm underneath it, and built a full Galaga emulator in VHDL that runs on FPGA hardware. I'm looking for full-time engineering roles starting after May 2027.

This site runs on Amplifly's idea. What kind of brain do you have today?

Protector 10 min

DND + Timer

You're already in it. The job now is defending it, so shut off the interruptions and put a hard edge on the session.

Simplified. The real recommender is a Thompson-sampling bandit that learns from your own logs instead of a lookup table.

Work

Five things I built, two of which you can poke at right here.

01 2024 — now

Amplifly

Founder and developer · App Store · amplifly.me

Amplifly asks one question in the morning, the one you just answered up top, and then hands you something specific to do about it. Most tools in this space give you general advice about wellness. This one gives you a technique matched to the state you're actually in.

The matching happens on your phone. It's a Thompson-sampling contextual bandit over twelve techniques, so it drifts toward what works for you rather than what works on average, and none of your data leaves the device. Fifteen neurodivergent profiles to start from, and you can add techniques of your own.

Five seconds. Pick the state you're in. That's the whole interaction, deliberately, so that it still happens on a foggy morning.

Amplifly Lite is the version that's live. It won first place in an industry investor pitch competition and a $7,000 grant along the way, and I was one of two students picked to present it to about 2,000 employees at Prudential Financial.

React Native
Expo
Supabase
Contextual bandits
App Store Connect

02 2026

NeuroStudy Coach

Deep learning · GitHub

A study planner for neurodivergent students, built around a network that predicts how long an assignment will actually take you, instead of asking you to estimate it yourself.

Prediction error against held-out sessions

Rule-based baseline

100%

Trained network

25%

A 75% reduction. Three layers, 128/64/32, batch norm and dropout, about 12,000 parameters, ~90 epochs on 2,000 generated sessions across eleven features.

Around the model sits a scheduling engine, a Streamlit interface, and a study assistant running on the Anthropic API. Everything stays on your machine.

PyTorch
Streamlit
Anthropic API
Python

03 Fall 2025

Galaga, in hardware

CPE 487 · VHDL on a NEXYS A7 · with Michael Moschello · GitHub

A fully playable Galaga emulator written entirely in VHDL, running on a Digilent Nexys A7-100T with no processor and no operating system underneath it. A custom VGA engine drives 800×600 at 60 Hz with no frame buffer, generating every pixel on the fly as the controller asks for it, and a finite state machine runs the whole game loop.

The fleet is a 6×10 formation of three enemy classes that breathes in and out while it tracks sideways, breaks off into curved dive attacks, and sends coordinated squad fly-ins from the edge of the screen. The starfield and the attack patterns come from a hash-based PRNG. Score goes to the seven-segment display, remaining lives to the board LEDs, and a results screen at the end reports shots fired, hits, and accuracy.

Ninety-eight seconds of the emulator running on the board, filmed off the monitor. Formation movement, dive attacks, enemy fire and the level counter.

A monitor showing the Galaga ready screen: a scrolling starfield, the green player ship, four red enemies and the text READY! with LEVEL 1 in the corner. — Ready screen, driven straight out of the FPGA over VGA

The results screen reading SHOTS FIRED 0024, NUMBER OF HITS 0017, HIT MISS RATIO 070%. — Ready screen, driven straight out of the FPGA over VGA

Browser demo, not the FPGA build SCORE 0 WAVE 1 SHIPS 3

This is not the VHDL project. It's a small JavaScript stand-in, here only so there's something to play in a browser. The real one is the hardware above: VHDL on an FPGA, no processor, no JavaScript, output over VGA. The repository has the VHDL source and the full write-up.

VHDL
Vivado
NEXYS A7 100T
Finite state machines

Cloud Analysis

Image processing and coding, final project · GitHub

A machine learning pipeline for cloud imagery. Preprocessing, feature extraction, classification, scored with cross-validation, and a look at the trade-off between model complexity and how well it generalized.

Python
OpenCV
scikit-learn

Movie recommendations

CPE 551 course project · GitHub

Scrapes IMDb for title, year, genre, rating and popularity, then recommends films against whatever filters you set. The ranking lives in a recommender class, and the notebook plots the distribution it's drawing on.

Python
pandas
Jupyter

Older things are on GitHub: Unity and Unreal games, a VR project, code for flying a DJI Tello drone. There's a portfolio PDF too.

Experience

Click any of them to open it up.

Built AI prototypes and ran rapid experiments to validate new product concepts and model-driven features. The kind of work where the question is still open and the cycles are short.

Worked with design and product on early concepts and turned them into things real users could actually sit down in front of.

Selected to teach a hands-on chemistry enrichment course to elementary students. Kodely runs K-12 programs across more than 100 schools and districts, reaching upwards of 20,000 students.

I grade for the embedded systems course, which covers microprocessor design, digital systems and the messy boundary between hardware and software, and I write detailed feedback on submissions.

On the admissions side I lead campus visits, including VIP ones, and I put together a Coffee Chat program so admitted students can talk to people who already go here before they decide.

Five summers teaching introductory programming, game design and creative computing to groups of ten to fifteen kids.

Skills

Hover a group to pull it forward.

Languages

Python
C++
VHDL
JavaScript
SQL

Machine learning

PyTorch
scikit-learn
pandas
NumPy
Prompt design
Model evaluation
REST integration

Tools

Git
Xcode
App Store Connect
Streamlit
OpenCV
MATLAB
Vivado
OMNeT++
SolidWorks
Claude
Cursor
Copilot

Hardware

Digital logic design
Microprocessor systems
FPGA development
Arduino
Circuit design

I work with Claude, Cursor and Copilot every day, and I check what they hand me before it goes anywhere near a user.

School

Stevens Institute of Technology, Hoboken.

B.E. Computer Engineering
Concentration in artificial intelligence

Graduating: May 2027
GPA: 3.988 / 4.0
Year: Senior

Coursework

Applied Machine Learning graduate level
Introduction to AI Engineering in progress
Computational Data Structures and Algorithms
Probability and Statistics with Data Science Applications
Modeling and Simulation
Digital System Design
Microprocessor Systems
Image Processing and Coding
Discrete Mathematics

Honors

Presidential Scholarship
Edwin A. Stevens Scholarship
Martha Bayard Stevens Scholarship
Tau Beta Pi engineering honor society
IEEE-Eta Kappa Nu
Gear and Triangle service honor society
iSTEM@Stevens program for innovators

Outside class

President, Order of Omega
Sergeant at arms, Khoda
Sigma Delta Tau, several roles
Student government senator, campus events

Let's talk.

I'm looking for full-time software and machine learning engineering roles starting after I graduate in May 2027. Email is the best way to reach me and I'll write back.

Email lilyaddy15@icloud.com LinkedIn lily-stone GitHub Lilyaddy15 Résumé PDF

LilyStone