End-to-End AI Projects in Biomedical Computing

I'm Arel, a high-school student teaching myself applied machine learning by building complete projects — mostly in medicine and biology. For each one I start from a public dataset, train and evaluate a model, and ship it as an app or an in-browser demo you can try. What I care about most is being honest about the results: every project reports real metrics on held-out data and says where it falls short. Everything here is educational — none of it is a medical device.

Educational content only. Nothing on this site constitutes medical, financial, or professional advice, and none of these projects are certified products or devices. See the full disclaimer.

Biomedical projects

MoleCheck

Privacy-preserving skin-lesion analysis: a YOLO11 computer-vision classifier trained on 11,720 dermatology images, running fully on-device via Flutter and TensorFlow Lite so no user data leaves the phone.

Computer vision TFLite Flutter Live demo

ECG Arrhythmia Classification

A 1D convolutional neural network for heartbeat classification on the MIT-BIH database, with a rigorous inter-patient evaluation documenting the drop from 96% in controlled settings to 44% on unseen patients.

Signal processing 1D CNN ONNX Live demo

ADHD Classification from EEG

A multi-channel 1D CNN classifying pediatric EEG by study group, evaluated with a strict subject-level split to measure genuine generalization — 92% accuracy on unseen subjects.

EEG 1D CNN ONNX Live demo

Leukemia Cell Subtyping

A computer-vision classifier subtyping leukemic blood cells across four classes at 99.8%, with a critical analysis of why that figure reflects a dataset limitation rather than true clinical performance.

Computer vision YOLO ONNX Live demo

Additional projects

Two projects outside the biomedical focus, included to show broader range in computer science — reinforcement learning and applied machine learning on financial time series.

Stock Volatility & Sentiment

A neural network predicting next-week volatility risk from price and news-sentiment features, including a controlled ablation demonstrating that the sentiment signal did not improve performance.

Fintech Time series Sentiment Live demo

Autonomous Pentest Agent

A reinforcement-learning agent that learns efficient vulnerability-discovery strategies against a sandboxed web application, outperforming a random baseline sixfold — with a defensive extension for detecting agentic ransomware.

Reinforcement learning Security Q-learning Live demo

Computer Science + Bioengineering

The biomedical projects above map onto the CS + Bioengineering path at Illinois: the EEG and ECG work sits alongside Grainger's neural-engineering focus, while the dermatology and blood-smear classifiers align with its bioimaging and computational-and-systems-biology research — all built with the honest, leakage-aware evaluation the field demands. Read the full CS + BIOE breakdown →