ut.Let's talk
All projects
Brain–computer interfaces / Academic prototype

Mindplay.

Three signals. One hands-free interaction.

Explore the repository
Illustrated Mindplay headset calibration from the supplied animated walkthrough
ANIMATED WORKFLOW ILLUSTRATION · NOT LIVE FOOTAGE
WHAT IT DOES

A team-built brain–computer interface prototype combining motor-imagery EEG, intentional blinks, and head motion into hands-free game controls.

MY CONTRIBUTION

Team project. The case study describes the system; individual task ownership is not yet specified.

THE INTERESTING DECISION

Motor imagery, blink, and head-motion detectors share a cooldown so simultaneous detections do not produce competing commands. Refractory periods and consecutive predictions address false activation.

START WITH THE EVIDENCE94.4% / 160 trials

Repository-reported accuracy and balanced accuracy for raw channels 0,3,4 cropped to 3 seconds. The report evaluates a saved model on a matching dataset; it does not establish a held-out split or across-user generalization.

Inspect the source ↗
BUILT WITH
PythonFBCSPLDALSLPyQt6pytest
PROJECT TYPE

Academic prototype

START HEREREADME & setup
01 / THE IDEA

Three signals. One hands-free interaction.

The problem

Explore an alternative to conventional game controllers for players with limited hand mobility. The engineering challenge is translating noisy, calibrated sensor signals into intentional commands without competing detectors causing accidental actions.

The approach

Smarting headset signals arrive through Lab Streaming Layer. A five-band FBCSP/LDA branch classifies motor imagery, while blink and gyroscope branches detect complementary inputs. Shared command arbitration and cooldown feed keyboard controls and a live gamepad overlay; a PyQt6 interface connects recording, training, evaluation, and live use.

02 / THE SYSTEM AT A GLANCE

Follow the flow.

  1. 01Headset / LSL
  2. 02EEG / blink / motion branches
  3. 03Command arbitration
  4. 04Shared cooldown
  5. 05Game controls + overlay

A simplified architecture overview. The repository documents the implementation details.

03 / THE ENGINEERING DECISIONS

Why it's built this way.

01

Arbitrate competing inputs

Motor imagery, blink, and head-motion detectors share a cooldown so simultaneous detections do not produce competing commands. Refractory periods and consecutive predictions address false activation.

02

Make units explicit

The pipeline expects microvolts. Stream values in volts require conversion, and the recording/training channel ordering must match.

03

Use subject-specific training

Record balanced motor-imagery and rest trials, then train and evaluate before live use. Separate matching-dataset results from proxy cross-dataset checks rather than presenting every score as subject accuracy.

04

Connect the workflow

The desktop interface brings recording, training, evaluation, and real-time use into a single flow. Low evaluation accuracy offers a retry or continue choice.

04 / WHAT TO LOOK AT

Evidence over adjectives.

VISUAL WALKTHROUGH / ANIMATED EXPLANATION

User-supplied animation explaining headset setup, calibration, and hands-free gaming. This is an illustrated concept walkthrough, not a recording of the running prototype or a validation of its accuracy.

Read the visual workflow
  1. Fit the EEG headset and prepare the signal stream.
  2. Calibrate the signals before issuing game commands.
  3. Combine motor imagery, intentional blinks, and head motion for hands-free interaction.
Open video in a separate tab
Documented evaluation94.4% / 160 trials

Repository-reported accuracy and balanced accuracy for raw channels 0,3,4 cropped to 3 seconds. The report evaluates a saved model on a matching dataset; it does not establish a held-out split or across-user generalization.

ClassifierFBCSP + LDA

Hand motor imagery versus rest.

Complementary inputsEEG + blink + motion

Three branches with shared command arbitration and cooldown.

Automated checks13 tests reported

README reports passing checks for signal processing, channel selection, cooldown, blink buffering, and gyroscope processing. These are software checks, not proof of end-to-end usability.

Capabilities and result snapshots follow the project's current README, reviewed October 11, 2026. Read the source

Evaluation report & methodologySystem architectureDemo execution guide
05 / THE HONEST PART

What I'd improve next.

Current boundaries

  • Academic prototype, not a clinically validated accessibility aid, medical device, or safety-critical control system.
  • Matching-dataset scores may include training data; the report does not document an independent held-out protocol. Small datasets and best-available proxy scores must not be treated as generalization evidence.
  • Live behavior depends on the user, session, electrode placement, and calibration. Windows, suitable EEG hardware, and an external LSL stream are required for the complete setup.
  • Channel order, preprocessing, and units must match the trained model.

The next iteration

Document a leakage-free, held-out evaluation across sessions and users, then measure false activations and end-to-end command latency during real gameplay. Clarify my individual responsibilities within the team.

Open my Software & AI CV
KEEP EXPLORING

ForgeQueue

Workflows that survive the unhappy path.