uncoil

uncoil landing page
uncoil logo

Inspiration

Every move tells a story. For an athlete, it’s the determination to return to the field after an injury. For an elder, it’s the courage to regain independence after years of chronic pain. But far too often, those stories are interrupted—not by a lack of effort, but by a lack of accessible, reliable guidance. In physical therapy and fitness, the line between recovery and reinjury is drawn by one key factor: proper form. Without it, progress slows, and the risk of setbacks becomes a constant shadow.

We built Uncoil because we’ve seen these struggles firsthand—athletes held back by preventable injuries, loved ones frustrated by slow recoveries, and individuals left to guess if their movements are doing more harm than good. Through purposeful guidance and customized feedback, Uncoil ensures every movement is intentional, empowering people to heal with confidence and precision.

By harnessing innovative software and machine learning, Uncoil reimagines physical therapy as a space where expertise meets accessibility. At the intersection of care and technology, Uncoil transforms every motion into progress—reshaping how we move, heal, and thrive.

What it does

It combines personalized routines, real-time corrections, and hands-free feedback to create a dynamic and engaging physical therapy experience tailored to each user’s needs.
We use a specially trained machine learning model that leverages a custom deep neural network to create individualized physical therapy and fitness routines based on user-provided data, including weight, height, age, goals, and any other specific needs.
With almost 20 different exercises to choose from, clear visual references, and concise instructions for each exercise in the routine, we allow users to avoid any confusion and guide them toward their fitness and healing goals
Using a state-of-the-art machine learning model that analyzes user's posture and movements in real-time through live video recording, ensuring accuracy and safety during each exercise.
Immediate, actionable feedback via text-to-speech, allowing users to make corrections and focus on their exercises without needing to check the screen.
Machine learning model that analyzes the user's pose in real-time and offers correction.

How we built it

We built and trained our own custom deep neural network in TensorFlow to generate personalized exercise routines tailored to each user's inputted data and goals. For real-time pose estimation and feedback, we integrated the YOLO v11 object detection model in PyTorch, to allow us to accurately analyze user movements and to ensure proper posture during exercises. Using real-time text-to-speech, we give users immediate feedback to any poses in the incorrect form, allowing instant fixes. We then used GenAI, specifically the CoHere API, in order to process the user inputs and generate clear descriptions of each exercise. We built the backend infrastructure in Python using FastAPI, and leveraged React and TypeScript on the frontend.

Challenges we ran into

We ran into some computer hardware issues when testing our project using the webcam camera. It would begin to overheat, causing the computer to suddenly freeze and render unusable for short periods of time.
When using the YOLO v11 model, we found we needed to further fine-tune it in order to better detect the poses with greater accuracy
During the design process, we tossed several different potential designs into the air before finally settling on one we all felt proud and satisfied by

Accomplishments that we're proud of

We were able to train a deep neural network from scratch in just 24 hours
One of our best and most beautiful frontend displays :)
Each individual in the group had their own role, allowing us to be far more productive as we could work in parallel with one another. At the very end, we were able to come back together during the integration portion, having a unified ending to our project

What we learned

It was our first time trying out FastAPI on the backend as opposed to Flask. We found that FastAPI is a great alternative as a backend infrastructure as we found the syntax to be quite clean and declarative, making our integration process seamless.
We discovered just how intricate deep neural networks and machine learning can be. Having to create a model within the time frame highlighted the importance of understanding the complexities of these models in a way that would not be possible in any other circumstances.
We were able to better analyze and understand computer vision models to detect and estimate human poses.

What's next for Uncoil

We would want to use a larger object detection model to better estimate poses with higher degrees of accuracy. Perhaps even using several models in tandem with one another could be an interesting and unique direction to investigate further.
We would try and implement a feature that highlights the exact part of the body that incurred a posture error. This way, users would be able to have a strong visual for any necessary alterations.
We could train a more advanced machine learning model that takes in a larger number of user data inputs. This would ensure users have an even more customized routine.