Sentinel

Inspiration

Inspired by the accelerating loss of species and the rise in poaching, we wanted to show how small, inexpensive devices like Arduinos can have outsized conservation impact when combined with creative engineering and AI.

What We Built

An embedded-sensor AI poaching-detection network that combines edge audio classification, Bluetooth sensing, satellite imagery, and a central dashboard:

Gunshot detection: Custom AI model (based on feature extraction from YAMNet) running on edge nodes to detect gunshots and trigger alerts.

Bluetooth detection: Nodes scan for nearby Bluetooth devices to help identify human presence in restricted areas.

Shot triangulation: Time-of-arrival inputs from multiple nodes are used to triangulate the gunshot location.

Satellite imagery: Supplementary area data obtained via the BrightData API to help prioritize monitoring and validate detections.

Alerting & dashboard: Nodes send messages wirelessly to a central computer that displays node status and locations on a Leaflet map GUI.

How We Built It

Custom Classification model: Performed feature extraction using a YAMNet-derived pipeline, then trained a compact classifier on labeled gunshot and ambient audio samples. Embedded deployment: Ported the model to run on the Arduino-compatible edge hardware (audio capture, preprocessing, inference). Networking: Implemented Wi‑Fi communication from nodes to the central server (MQTT/HTTP-based messages). Central system: Built a GUI with Leaflet to show nodes and detections; implemented backend services to receive alerts, store events, and run triangulation. Satellite integration: Used the BrightData API to fetch satellite tiles/imagery for regions of interest and display them in the dashboard.

What We Learned

Practical experience deploying ML on constrained edge hardware: trade-offs between model size, latency, and accuracy. Real-world networking is messy: pairing many low-power nodes, reliability over Wi‑Fi, and message ordering all matter. Data engineering challenges: labeling audio, handling noisy real-world recordings, and integrating external APIs like BrightData. Rapid prototyping workflow improvements: modular code, smaller repos, and CI helped reduce friction as beginners.

Challenges Faced

Wi‑Fi communication from nodes to the central dashboard was harder than expected due to initial database and messaging design choices. Managing a large repository and unfamiliar toolchains slowed development. Collecting and labeling diverse gunshot audio samples for robust detection. Synchronizing timestamps for accurate triangulation across nodes with variable network delays.

Next Steps and Future Ideas!

Improve time synchronization (e.g., GPS or precision time protocol) to tighten triangulation accuracy. Optimize the edge model further for lower power and better false-positive control. Add secure, resilient mesh networking and encrypted messaging. Expand satellite analytics for automated hotspot detection.