The curve on the laptop is real. A small neural net (two inputs, four hidden units, one output) is learning XOR in your browser, and that line is its loss.
I’m building Sienna AI, an AI employee for small businesses.
Cornell University AutoBoat Project Team
Software Engineer to now
I write perception software for Cornell's autonomous boat. I added quaternion-derived yaw to the exponential moving averages used for vessel tracking, and buoy color and sign validation, so the autonomy stack gets more consistent information about tracked objects.
Electrical Systems to now
Contributed to Robotics Power Distribution + Binary Relay Board
Weill Cornell Medicine
ML Research Intern to
I built and evaluated deep learning pipelines that segment abdominal organs in MRI scans for radiotherapy planning. I worked on the dataset infrastructure, preprocessing and GPU training in Python, PyTorch and Linux, across hundreds of experiments. The segmentation system reached about 0.94 Dice.
Syracuse University College of Engineering and Computer Science
Robotics Systems Engineer to
I worked on a radar-equipped hexacopter built on Pixhawk, ArduPilot and a Raspberry Pi, on sensor integration and synthetic aperture radar processing. I wrote image-reconstruction workflows that use motion compensation and backprojection to account for the drone moving while the radar records.
SUNY Oswego
Computational Research Intern to
I wrote Python tools to analyze experimental data for research on ladder-type electromagnetically induced transparency in cesium. They were used to study the system’s optical response and to help interpret the measurements.
Things I've made
FIFA World Cup predictor
It turns historical international football results into match predictions and tournament forecasts. The hard part was draws. My first models reached about 53% accuracy by never predicting one: zero draws called across 8,155 test matches. So I stopped scoring on accuracy, trained with class-balanced weights, and judged every model by macro F1 on a chronological split, where ignoring an outcome costs a third of the score. Adding leakage-safe Elo ratings then lifted macro F1 about 18%, from a 0.445 baseline to 0.526, with a model that predicts all three outcomes. Knockout games can’t end level, so the bracket simulator splits each draw probability between the two teams before a Streamlit app runs 10,000 Monte Carlo tournaments to estimate every team’s chance of reaching each round.
Ask it a tactics question in plain language and it animates the answer on an editable 22-player pitch. The hard part was the simulation engine behind the model’s structured output: it models how defenders react, checks passing lanes and interceptions, and searches for supporting runs. Playback is deterministic, with pause, scrubbing and replay, plus opponent-response comparisons and undo/redo. It still works in full when the AI service is unavailable.
An iOS app that shows you how you spend your time, through periodic check-ins, an activity log and daily productivity charts. The hard part was the reminder scheduling: it reads the calendar and skips quiet hours and the events you pick. Notification actions open an activity entry that is already filled in.
Data Structures and OOP, Functional Programming, Discrete Math, Data Science and Decision Making, Linear Algebra, Differential Equations, Circuits
Activities
Autonomous Boat Project Team (Software), Institute of Electrical and Electronics Engineers (PR Team), South Asian Council (Publicity Team), Society of Asian Scientists and Engineers, AI Alignment, Muslim Educational & Cultural Association, Notion Campus Leader, Adobe Campus Ambassador
About me
I'm Rayhan, a computer science student at Cornell University interested in intelligent systems, with a focus on machine learning, autonomy, and data engineering. My experience spans software engineering, robotics, and applied machine learning.