Signal processing and ML pipeline for classifying attention and emotional state from raw EEG.

Bandpass filtering isolates relevant brainwave frequencies, and Independent Component Analysis (ICA) removes artifacts like eye blinks and muscle movement. Band-power features are extracted and fed into SVM and Random Forest classifiers. Drag the slider above to see how signal amplitude relates to the predicted state.

Python · MNE · scikit-learn · NumPy

A PPO agent learning continuous-action control for a simulated robotic arm.

Trained in a PyBullet and Gymnasium environment on a reach-and-grasp task using Stable-Baselines3. The demo above is simple inverse kinematics — dragging moves the end effector the way a human would. The actual RL agent learns this kind of motion on its own through trial and reward, without being told the geometry directly.

Python · PyBullet · Gymnasium · Stable-Baselines3

A full-stack system streaming live model predictions to an interactive dashboard.

A Python inference service serves predictions from a trained biosignal model, streamed to a Spring Boot backend via WebFlux and Server-Sent Events, and rendered on a front-end dashboard that updates live as data streams in. The demo above simulates that live stream.

Python · FastAPI · Spring Boot · WebFlux · JavaScript

Decoding a signal and using it to drive a robotic arm in real time — the full decode-to-action loop.

Connects the EEG classifier's signal-decoding approach with the robotic arm's inverse kinematics into a single pipeline: a simulated intent signal is decoded with a nearest-centroid classifier, and the recovered intent drives the arm toward the corresponding target in real time. The same decode-then-act shape as a real BCI-controlled prosthetic or robot, just with simulated signal in place of real neural data for now.

Python · NumPy · Matplotlib · JavaScript