The problem AlignPose 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.

Challenges I 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.