AI voice assistance for runners

We want to make an app can integrate with HealthKit on Apple (our current target) and support runners who wears earbuds or headphones while running to ensure their health is preserved.

Comment

Inspiration

  • PulseRunner was inspired by the need to bridge the gap between raw sensor data and elite-level coaching for everyday runners.
  • Traditional marathon training relies on static, linear periodization models—rigid schedules that prescribe mileage weeks in advance without considering the runner's physiological state.
  • Research indicates that increasing run distance by just 10%–30% in a single session increases injury risk by 64%.
  • With annual injury rates estimated between 37% and 56%, generic plans often fail to detect "invisible" stressors like poor sleep or illness that reduce load tolerance.

What it does

  • PulseRunner is an AI coaching application that collects high-frequency streaming biometrics, including heart rate, pace, and GPS data.
  • It provides immediate audio feedback during runs for safety and automated, weekly adjustments to training schedules based on recovery trends.
  • The system utilizes a dual-layer approach to provide recommendations before, during, and after running activities.
  • It detects signs of non-functional overreaching, such as an elevated resting heart rate, to adjust training loads before injury occurs.

How we built it

  • We implemented a Lambda Architecture to balance the requirements for low-latency safety alerts and high-accuracy historical analysis.
  • Data Ingestion: Amazon Kinesis acts as a high-throughput buffer for incoming data, decoupling ingestion from processing.
  • Speed Layer: Apache Flink performs stateful computations over data streams to detect anomalies and trigger pacing alerts in real-time.
  • Batch Layer: Apache Airflow orchestrates weekly plan updates, running ML inference on historical trends stored in an S3 Data Lake.
  • Polyglot Persistence: We used InfluxDB for high-write sensor data and PostgreSQL for structured user profiles and training plans.
  • Edge Processing: A React Native mobile app aggregates 1 Hz sensor data into compressed micro-batches to preserve battery life.

Challenges we ran into

  • Traffic Spikes: Running telemetry is "bursty," with massive traffic spikes during peak hours (e.g., Saturday mornings) that could crash traditional databases.
  • Connectivity Issues: Data often arrives out of order due to poor cellular connectivity; we used Apache Flink to re-order late-arriving events based on "Event Time."
  • Write Amplification: Storing millions of heart rate records in a standard relational database would degrade index performance.
  • Energy Management: Streaming every heartbeat individually is inefficient, necessitating a "Store-and-Forward" strategy on the mobile device to reduce network overhead by 98%.

Accomplishments that we're proud of

  • Real-Time Safety Interventions: Built a logic flow that evaluates conditions instantly: $$current_heart_rate > (max_hr_threshold \times 0.95) \text{ for } > 30 \text{ seconds}$$ If met, the system publishes immediate "Slow Down" warnings via SNS.
  • Adaptive Recovery Logic: Successfully implemented a health check that compares $acute_load$ (last 7 days) against $chronic_load$ (last 28 days): $$\text{IF } acute_load > (chronic_load \times 1.3) \implies \text{Status: "OVERTRAINING_RISK"}$$ This triggers a 10% reduction in the next week's mileage.
  • Scalable Infrastructure: Deployed a partitioned Kinesis stream that distributes load by $user_id$ and scales automatically based on throughput.

What we learned

  • Storage Optimization: Purpose-built time-series databases like Amazon Timestream can compress sensor data by over 90% and automate downsampling.
  • Architectural Trade-offs: A single pipeline cannot efficiently serve both sub-second safety alerts and deep historical analysis, confirming the necessity of bifurcating the data flow.
  • Schema Evolution: Using JSONB in PostgreSQL for training schedules allowed us to iterate on coaching logic without requiring costly database migrations.

What's next for AI voice assistance for runners

  • IoT Sensor Integration: Integrating with Bluetooth-enabled "Smart Shoe" insoles to track "Ground Contact Time" and "Impact Force" to detect form degradation before pain begins.
  • Computer Vision Form Analysis: Allowing users to upload a 10-second video of their stride for analysis of biomechanics like hip drop or overstriding.
  • Social "Ghost Runner" Mode: Utilizing the low-latency Speed Layer to provide real-time audio updates about a user's position relative to a friend in a different location.
Share this project:

Updates