Student developer and researcher · Greenwich, Connecticut
I build software and AI tools to solve problems I can test against real data. Since 8th grade I have completed four research and engineering projects, each with larger, less pre-packaged data than the last. My most recent, Tailwater, is a flood-risk tool that the Greenwich Department of Public Works is using.
I have been doing research for four years and I am still going. What keeps me at it is wanting to help my community, whether that means a flood tool for my town or something I can teach in person.
That is why I volunteer to teach AI literacy. I have worked with residents at senior living homes, showing them how scammers use AI to clone the voice of a grandchild and call to ask for money, and how to protect themselves from those calls.
Teaching residents at a senior living home how to spot AI voice-cloning scams
I have also been entrepreneurial since I was a kid. In 3rd grade I gave a speech about my mom's kitchen for Junior Achievement, and I have enjoyed turning ideas into real things ever since. Outside of school and research, I run cross country and track and field.
Projects
Grade 8 · Middle School STEM Fair, May 2024
Student Support Chatbot
My school had one counselor for the entire 8th grade, and students had questions outside school hours. I built a chatbot that answers them from the counselor's own materials, available any time.
The counselor gave me FAQs and documents on moving between grades, the student portal, and the pattern schedule. I split them into searchable chunks, matched each question to the closest ones, and used a language model to write the answer.
Tools: large language model, text embeddings, vector database, open-source NLP models.
Aircraft drop fire retardant mostly by pilot judgment, which wastes chemicals and can miss the fire. I built a model that predicts fire spread and recommends drop zones from terrain, weather and satellite data, and I checked it against the January 2025 Palisades Fire.
It uses USGS LiDAR elevation data, NOAA weather records, and NASA MODIS and FIRMS fire data, with 3D terrain simulations in Python. The model reached about 85 to 90 percent accuracy in my tests.
Tools: Python, PyVista, neural networks, linear regression, satellite and LiDAR data.
Limitations: this work was done in simulation. It has not been tested on a live fire.
A Paper Test Strip for Parkinson's-Linked Compounds
Parkinson's is usually diagnosed after most of the affected neurons are already lost. Studies show that certain airborne compounds in earwax differ in patients, so I built a low-cost color-changing strip and an AI that reads it from a phone photo.
I made synthetic earwax, spiked it with pentanal at known concentrations, and exposed reagent-treated strips to the vapor. A computer vision step finds the strip and measures its color, regression models estimate the concentration, and an iPhone app shows the result with a confidence score. In controlled testing on blinded samples it classified about 90 percent correctly.
Tools: OpenCV, NumPy, ridge and logistic regression, cross-validation, an iPhone app, wet-lab chemistry.
Limitations: this is a proof of concept on synthetic samples. It needs more calibration and clinical validation before it could be used to screen people.
Finding the storm drains most likely to flood normally takes an engineering study. Greenwich paid $125,000 for one that covered 18 locations, and the town has about 6,000 catch basins. Tailwater scores every basin before a storm using data the town already has.
A physics model sets each basin's risk: Manning's equation for pipe capacity, the Rational Method for storm runoff, and NOAA Atlas 14 rainfall, with an added penalty where the brook backs water up into the pipes. A separate machine learning model trained on over 13,000 simulated storms gives a flood probability alongside it, scoring 0.95 AUC on basins it had not seen. The dashboard shows a live risk map from the NOAA forecast, a full calculation for each basin, a plain-English field report written with the Claude API, and exports DPW uses for its stormwater permit and capital planning.
My first idea was a sensor in every drain. Greenwich engineers showed me it would miss the real cause of flooding, so I changed direction. I met with them eight times between March and October 2026, and each meeting changed the app. It now also scores a second neighborhood that has no engineering study, and I am working on expanding it to other towns.
Tools: Python back end, scikit-learn, interactive map, NOAA and FEMA data, Claude API.
Limitations: Tailwater is a screening tool, not an engineer's stamped analysis. Pipe slope, roughness and runoff use assumed values, and most of the machine learning training data is simulated.
I am testing whether the branching tunnels of a termite mound can move heat away from servers with less air resistance than straight fins. The plan is to 3D print the channel shapes, cast them in ceramic using the lost-PLA method, and measure airflow in a wind tunnel.
Skills by year
Grade 8
Language models, embeddings and vector search, building from a client's own documents.
Grade 9
Neural networks, geospatial and satellite data, 3D simulation, validating a model against a real event.
Grade 10
Computer vision, calibration and blinded testing, lab chemistry, mobile app development.
Grade 11
Engineering formulas in code, fair train and test splits, full-stack web development, working with professional engineers.
