This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass
What I Built
OpenStride is a completely offline, privacy-first wearable running coach.
Modern running has become obsessed with screens. We stare at smartwatches for our heart rate, check pacing on our phones, and get distracted by Strava notifications mid-stride. OpenStride removes the screen from the run entirely.
You clip a simple ESP32 + MPU6050 sensor to your shoe or waistband. It silently logs your running kinematics (cadence, impact force, and leg fatigue) to a MicroSD card while you are out on the trail. You run completely disconnected—no phone, no internet, no distractions.
When you get home, your local PC running Google's open-weight Gemma 2 model (via Ollama) ingests the raw data. It analyzes your form, calculates fatigue drops, asks how you felt, and generates a customized marathon training plan for tomorrow. The screen is only used for two minutes after you've touched grass.
Demo
https://github.com/dinesh9-ai/Hacktoberfest2026_withAI/blob/main/DevChallenge2/Demo/Demo_video.mp4
Code
https://github.com/dinesh9-ai/Hacktoberfest2026_withAI/tree/main/DevChallenge2/Code
How I Built It
- OpenStride is built entirely around an open-source hardware and local AI architecture:
- Hardware: An ESP32 microcontroller paired with an MPU6050 **accelerometer handles the edge data logging. It is lightweight, cheap, and **requires no internet connection.
- Data Pipeline: Python and Pandas process the raw CSV data into readable biomechanical metrics (like impact magnitude and fatigue ratios).
- The AI Core: I used Gemma 2, Google's powerful open-weight model, running entirely locally on my PC via Ollama. Gemma acts as the reasoning engine, taking the cold, hard math from the pandas dataframe and translating it into empathetic, actionable coaching advice.
Why Does Open Innovation Matter?
Open innovation isn't just a feature for this project; it is the only way it works.
-** Absolute Privacy:** Running kinematics and location timings are highly personal biometric data. By using a local open-weight model (Gemma) instead of a closed cloud API, the runner's health status and routines never leave their laptop.
- True "Touch Grass" Capability: Marathoners often train in remote locations, national parks, or mountain trails with zero cell service. Because OpenStride relies on local hardware and local PC inference, you could be in a remote cabin for a high-altitude training camp with no Wi-Fi, and your AI coach still works perfectly.
- Zero Ongoing Costs: Daily coaching via a proprietary LLM API would rack up costs quickly. Running Gemma locally means runners can query their coach, iterate on their goals, and analyze endless historical runs for absolutely free.
My Agent Session
NA
Prize Categories
- Best Use of Gemma: I utilized Gemma 2, Google's open-weight model, running locally via Ollama to serve as the project's core AI reasoning engine, eliminating the need for cloud APIs.
- Best Use of Entire: I have included my agent session in the write-up above to share how the underlying data pipelines were built.
Top comments (0)