BodyTalk

Candidates rehearse their answers but rarely get specific feedback on nonverbal behaviour, and uploading a personal interview recording to a cloud service is a privacy risk.

BodyTalk runs the whole analysis locally and never touches the network while it runs. Its feedback describes only what was visible, never how someone felt or how they would fare in a real interview.

• 01On-device pipeline: frame extraction, then MediaPipe Holistic for face, pose and hand landmarks
• 02Three parallel channel analysers combined by confidence-weighted adaptive fusion, the project's research contribution
• 03A fixed-weight fusion baseline behind the same interface, so both methods can be compared on one video
• 04Interactive dashboard with timestamped insights and a weight-over-time chart
• 05PDF session report and session history stored in SQLite
• 06Sandboxed renderer that reaches the backend only through a preload bridge
• 07Windows installer, with 12 automated test suites across Vitest and pytest