Here’s a fact that sounds like science fiction but is just physics: the Wi-Fi signal filling your home is constantly being distorted by everything it passes through and bounces off — walls, furniture, and people. If you can read those distortions precisely enough, you can detect where a person is, and even roughly what pose they’re in, without a camera. It’s called Wi-Fi sensing, and the raw material is something called Channel State Information (CSI).

I wanted to see it work with my own hands, on cheap hardware, so I built a sensing node.

The build

The sensor is an ESP32-S3 — a microcontroller that costs about the price of a coffee. Flashed with open-source CSI firmware, it does one job very well: for every Wi-Fi packet it hears, it captures the fine-grained channel measurements and streams them out over the network to a machine that does the heavy analysis. On the receiving side runs an open-source Wi-Fi “DensePose” pipeline in a container, turning that stream of raw radio measurements into a live view of what the signal is telling us about the space.

End to end, it works: the node pushes CSI frames at around 46 per second, the dashboard picks them up, and it flips from its fallback simulation mode to “real hardware connected.” Watching a graph respond to someone moving in a room it can’t see is genuinely uncanny.

The bug that taught me to read the fallback logic

The first time I wired it up, the dashboard insisted it was in simulation mode even though my hardware was clearly streaming real data. I could see the packets leaving the ESP32.

The cause was a subtle one worth remembering. The software had an “auto-detect” mode that waited a short window to decide whether real hardware was present — and my ESP32’s first packet arrived just after that window closed. So it gave up, declared “no hardware,” and fell back to simulated data, ignoring the real stream that showed up a moment later. Forcing it to expect a hardware source instead of auto-detecting fixed it instantly.

The lesson: when a system has an automatic “is the real thing there?” check and it guesses wrong, don’t fight the symptom — find the timing assumption behind the guess. Auto-detection is convenience that quietly makes decisions for you, and when it’s wrong it’s wrong silently.

Why this matters

Wi-Fi sensing is a real and slightly unsettling capability: presence and motion detection that needs no camera, works in the dark, and sees through walls, built from hardware anyone can buy. For a security-minded person that cuts both ways — it’s a privacy consideration worth understanding, and it’s a fascinating, low-cost way to get hands-on with the physical layer of the networks we otherwise treat as invisible plumbing.

Tech: ESP32-S3, Wi-Fi CSI, ESP-IDF, Docker, an open-source Wi-Fi DensePose pipeline.