FitGuru
Building a Private, On-Device AI Workout Coach.
Created on 10th October 2026
•
FitGuru
Building a Private, On-Device AI Workout Coach.
The problem FitGuru solves
The Problem It Solves
Working out without guidance often leads to injuries from poor form, while personal trainers are expensive and inaccessible to most. Traditional fitness apps only count reps blindly, force users to tap sweaty screens mid-set, or compromise privacy by uploading live video feeds to the cloud.
What People Use It For & How It Helps:
Prevent Injuries with Real - Time Form Correction: Tracks body joint angles at 30+ FPS entirely on-device (via Vision AI) and alerts you immediately if your back curves or knees cave in - before bad posture causes pain or injury.
100% Hands-Free Workouts: Users control their entire session by voice ("Start push-ups", "Finish set", "Start rest") and Siri, eliminating the need to ever touch the screen with chalky or sweaty hands.
Guaranteed Privacy: Camera frames and workout data are processed 100% locally on your phone - no video or biometric data is ever sent to external cloud servers.
Smart Rest & Recovery: Automatically times rest intervals and gives contextual audio feedback on rep quality, making solo training as structured and motivating as having an elite personal coach in your pocket.
Challenges we ran into
Audio Clashes & Dropped Microphone Sessions
Hurdle: When the AI coach spoke aloud to give cues, iOS audio ducking killed the microphone session or caused the mic to hear its own voice, turning listening off mid-set.
Solution:Built an audio state machine synced with playback delegates (FitGuru.TTS.Finished) to cleanly pause listening during speech and instantly re-arm the mic with fresh buffers the millisecond audio finishes.
Vision Jitter & False Rep Triggers
Hurdle: Processing poses at 30+ FPS produced noisy joint angles, causing occasional double-counting or sets completing prematurely.
Solution :Created a mathematical Reflex Engine with angle hysteresis and debounce windows, requiring sustained geometric transitions before locking in a valid rep.
Flexible Hands-Free Voice Commands
Hurdle: Standard voice shortcuts forced rigid phrasing and often required awkwardly repeating the app's name mid-workout.
Solution: Paired Apple’s App Intents with fuzzy intent matching, allowing natural variations ("Finish set", "End set", "Stop set") to trigger instantly hands-free.
Tracks Applied (4)
Overall Winning Team
ElevenLabs
Best Use of ElevenLabs
Major League Hacking
Best Use of Gemini API
Major League Hacking
Best use of Agentic AI / Gen AI
Technologies used
Cheer Project
Cheering for a project means supporting a project you like with as little as 0.0025 ETH. Right now, you can Cheer using ETH on Arbitrum, Optimism and Base.


