FormForge

FormForge turns exercise analysis into something people can immediately understand. It shows a user’s original movement beside the same recording with its form corrected, combining deterministic pose geometry, a live angle meter, multilingual AI coaching, and frame-timed voice guidance. The hackathon proof of concept focuses on one clear and testable case: a side-view high plank in which the hips leave the shoulder–hip–ankle line. The original panel displays the measured body path, changing from red toward green as the position approaches the configured target. The corrected panel keeps the original video visible while drawing the corrected pose as a cyan hologram. A live angle meter compares the current position with the target, with a straight shoulder–hip–ankle line displayed as 180 degrees. The user can also open a report containing the hold duration, peak deviation, time outside the target band, worst frame, and coaching cues. MiniCPM-V-4.7-1B first examines sampled frames and identifies the movement. If it is not confident, DeepSeek-V4-Flash-Vision-Exp is used as a fallback. Both are accessed through AMD Token Factory. MediaPipe Pose then extracts the joint positions. Deterministic Python selects the most visible side, measures the hip’s signed angular offset and perpendicular distance from the shoulder–ankle line, detects the misalignment, and produces corrected keypoints. The correction is verified before it is displayed. Gemma-4-31B-IT receives only the validated finding and its frame span, then writes two or three concise coaching cues in the user’s selected language. Gradium converts those completed cues into speech, with each line triggered as the video reaches the relevant frames. The result is an experience designed for understanding rather than scoring. Instead of merely telling someone that their position is wrong, FormForge lets them see the difference on their own body and connect the visual correction with a simple spoken instruction.