AlignPose
Align your spine. Align your focus.
About
The problem it solves **AlignPose** helps students and professionals correct bad sitting posture before it turns into chronic pain. Using real-time AI pose detection directly in the browser, it delivers accurate, private, and non-intrusive posture feedback without any additional hardware.
How it's made
Challenges we ran into **1. Achieving Real-Time Performance in the Browser** Running pose detection models directly in the browser caused noticeable lag and high CPU usage, especially on low-end laptops. This affected both frame rate and user experience. *How I solved it:* I optimized the pipeline by reducing the input resolution, limiting the frame processing rate, and selecting a lightweight TensorFlow.js pose model. I also used requestAnimationFrame and efficient canvas rendering to keep the UI responsive while maintaining acceptable detection accuracy. **2. Defining “Good Posture” Accurately** Human posture varies widely between individuals, and a fixed “ideal posture” model produced many false alerts for users with different body proportions and seating styles. *How I solved it:* I implemented a calibration step where the user sits straight once to record a personal baseline posture. All future posture checks are compared against this baseline using joint angles instead of raw coordinates, making detection more accurate and personalized. **3. Reducing False Positives from Small Movements** Small natural movements like breathing, typing, or shifting in the chair were sometimes detected as posture violations, causing unnecessary alerts. *How I solved it:* I added temporal smoothing and threshold-based logic so that posture is flagged only if incorrect alignment persists for several consecutive frames. This reduced noise and made feedback more stable and less distracting. **4. Maintaining User Privacy** Most webcam-based systems upload frames to servers for processing, which raises serious privacy concerns. *How I solved it:* I designed AlignPose to run entirely client-side using TensorFlow.js. No images or video frames are stored or transmitted. All posture analysis happens locally in the browser, ensuring complete data privacy. **5. Integrating Visual Feedback Without Being Intrusive** Early versions displayed constant warnings that distracted users from their work. *How I solved it:* I redesigned the feedback system to trigger alerts only after sustained poor posture and used subtle visual cues instead of aggressive pop-ups. This improved usability while still encouraging correction.