AngelWare

Real-time inference for search and rescue: WiFi sensing and computer vision to find people where firefighters cannot see.

AngelWare pairs an ESP32 mesh that reads WiFi channel state information with computer vision models to locate survivors in zero-visibility conditions. A Raspberry Pi touchscreen streams frames to GPU inference in the cloud, and a voice agent calls ahead to the hospital.

Problem

Smoke takes away the one sense rescuers rely on, and radios are hard to use mid-rescue. The system had to detect presence without cameras alone, run pose estimation fast enough to be useful, and dispatch information where a wrong output carries real cost.

Process

CSI data is z-score normalized per subcarrier against an empty-room baseline to isolate human motion from RF noise, and RSSI is smoothed into near, medium, and far confidence zones. Frames travel over a binary WebSocket protocol as JPEG, with no JSON on the hot path, to a Detectron2 and DensePose server on RunPod GPUs. Bad frames are logged and skipped instead of dropping the stream.

Solution

• ESP32 CSI sensing with per-subcarrier feature extraction and RSSI proximity modeling.
• Edge-to-cloud streaming: Pygame touchscreen client on a Raspberry Pi, PyTorch DensePose inference on RunPod.
• ElevenLabs conversational voice and Twilio calls for autonomous dispatch to hospitals.
• MongoDB Atlas schema logging sessions, motion events, and sensor samples for future model training.

Results

• Real-time Pose inference on cloud GPUs
• 3 Signals: CSI, RSSI, vision
• 0 JSON on the frame hot path
• 1 Weekend, LA Hacks 2026

Justin Moran · Software Engineer · Chicago, IL · LinkedIn · GitHub · Resume (PDF)