Body Language Tips / About the tool

How the practice mirror works — method and limits

Updated 1 October 2026 · Body Language Tips

This is a camera-setup and self-practice tool. It uses visible landmarks to describe framing and changes during an attempt. It is not a psychological or interview-performance assessment.

Local processing

The browser downloads pinned MediaPipe face and pose models and a runtime, then runs them in a background worker. Camera frames are processed in memory. No microphone is requested and no recording is created. There is no server AI call or generated transcript.

Face tracked

The share of practice time with exactly one detected face whose key landmarks are inside the image and sufficiently large. Multiple faces, a cropped face or missing landmarks are rejected. A missed frame is not evidence of a particular behavior.

Shoulders in frame

Both shoulder landmarks must be inside the image with model visibility and presence above 0.65. The displayed percentage uses face-tracked time as its denominator. This describes the camera crop, not uprightness, health or good posture.

Near starting direction

Before the attempt, the tool takes the median of recent usable face observations. It compares the nose's horizontal offset from the midpoint of the outer eye landmarks, normalized by their distance, and the slope of that eye line. Frames within an offset difference of 0.12 and slope difference of 0.15 radians count as near the starting direction.

These are heuristic thresholds, not validated head-angle or eye-contact measurements. Eye position, facial anatomy, camera lens, occlusion and movement can affect them. The starting position is your reference, not a prescribed ideal.

Time and tracking quality

The tool attempts roughly six frames per second. Each sample represents no more than half a second; gaps are treated as missing data. Head-direction and shoulder percentages are withheld if fewer than five seconds are face-tracked or less than half the attempt is analyzed. The movement plot shows face-center displacement relative to the starting eye-landmark distance; it has no pass/fail threshold.

Limits and intended use

Lighting, glasses, occlusion, hardware, camera angle and person-to-person differences can change tracking. The tool has not been validated across demographic groups or accessibility needs. Use it for voluntary personal rehearsal, not evaluating people for hiring or another consequential decision.

It does not infer confidence, emotion, honesty, personality or competence. It does not identify people, classify hand gestures, analyze speech or predict what an interviewer thinks. No measurement is a target you need to maximize.

Technical references

• Google Face Landmarker for Web
• Google Pose Landmarker for Web
• MediaPipe local processing documentation

Runtime: @mediapipe/tasks-vision 1.0.1; version 1 face and lite-pose assets. MediaPipe components are distributed with Apache 2.0 attribution. See license.