Spottair | Devpost

Inspiration

The Quantity Trap:

Modern fitness apps have a fatal flaw: They gamify volume, not value.

The "Streak" Epidemic: Apps reward users for "hitting 100 pushups" or "closing rings." This encourages users to cheat on form just to hit the number.
Garbage In, Injury Out: Doing 50 squats with collapsed knees is significantly worse than doing 10 perfect ones. Yet, standard trackers treat them exactly the same.
The "Invisible" Danger: A user might feel productive because they logged a workout, but they are unknowingly reinforcing poor biomechanical patterns that lead to chronic injury.

The Spark:

We realized that how you train matters infinitely more than how much you train. Spottair make you to stop counting "reps" and start grading "movement quality." Spottair is an AI agent cares more about your spine angle than your streak!

What it does

Spottair is an AI Biomechanics Expert that shifts the focus from Volume to Technique.

Quality Over Quantity: While other apps celebrate that you did 10 reps, Spottair might tell you, "Only 6 of those counted. The last 4 lacked depth." It enforces a standard of excellence.
Zero-Friction Context: You don't enter data. You just move. Spottair observes your session and logs not just the numbers, but the integrity of the session.
Injury Prevention Protocol: It identifies "Form Fatigue" - the exact moment your technique breaks down - and advises you to stop, preventing "junk volume" training. Your rep count increases only when you have fully completed one set.

How we built it

The Vision Layer (Data Ingestion): We use MediaPipe to extract 33 skeletal landmarks in real-time.
The Classification Engine (The Brain): We trained a custom classification model using MLP on a dataset of labeled exercises.
The Agentic Feedback: While the classifier detects what went wrong, the LLM Agent explains how to fix it. It translates the model's classification label and extracted landmarks angles into natural language coaching - "Push your left knee outward"

Challenges we ran into

Data Imbalance: Collecting data for "Good Form" was easy, but getting enough samples of specific "Bad Form" variations to train the model was difficult. We had to simulate various errors to balance the dataset.
Defining "Perfect": "Good form" is subjective and varies by body type. We had to train the model to generalize across different body proportions (femur length vs. torso length) rather than overfitting to one body type.
Context Window Management: Feeding "every single frame" of a workout into an LLM for the summary is too expensive. We had to implement a summarization algorithm that identifies key moments of interest and set statistics to send to the LLM for analysis.

Accomplishments that we're proud of

High-Accuracy Classification: Achieving a high F1-score on our custom test set, successfully distinguishing between a "Good Squat" and a "Bad Squat".
Invisible Logging: The user completes a workout without touching their phone once, yet ends up with a dataset richer than any manual log. -The "Coach's Reply": Successfully generating a post-workout summary with reply video for each set.

What we learned

Model Context Matters: A trained model is superior to geometric rules because it learns the temporal dynamics of a movement, not just static positions.
Feedback Depth: Real-time feedback fixes the rep, but post-set feedback fixes the habit. Both are necessary for true progress.

What's next for Spottair

Physiotherapy Integration: Exporting the "movement quality" reports to doctors to help rehab patients recover from injury.
Multi-Angle Support: Using two phones simultaneously to get a 3D analysis of the user.