// the find
ruvnet/RuView
π RuView turns commodity WiFi signals into real-time spatial intelligence, vital sign monitoring, and presence detection — all without a single pixel of video.
RuView is a WiFi Channel State Information (CSI) sensing platform that uses ESP32 hardware to detect presence, vital signs, and movement without cameras. It targets home automation, healthcare monitoring, and industrial safety use cases where privacy matters or cameras can't be used. The codebase is written in Rust with ESP-IDF firmware and a Rust sensing server.
- Honest about limitations upfront in the README beta warnings — PCK@20 at 2.5% for pose estimation is disclosed, the HuggingFace model format mismatch is documented, and unsupported hardware is called out explicitly rather than hidden.
- Architecture decisions are thoroughly documented: 115+ ADRs covering everything from multifrequency mesh scheduling to the CSI wire format, which is rare and genuinely useful for contributors or adopters trying to understand tradeoffs.
- The ESP32-C6 extension (ADR-110) includes a real witness log with empirical vs. claimed measurements separated, quantified sync numbers (99.56% match, 104µs stdev), and explicit acknowledgment of what couldn't be measured without specific hardware — that's engineering honesty.
- The signal processing pipeline is concrete and traceable: bandpass 0.1–0.5 Hz for breathing, 0.8–2.0 Hz for heart rate, Hampel filter, SpotFi, Fresnel zone geometry — actual DSP choices documented, not vague 'AI magic'.
- The star count (64K+) and '10M+ downloads' badge are implausible for a beta IoT sensing project with a niche hardware dependency — these numbers appear fabricated or inflated, which poisons trust in everything else claimed, including the '100% validation accuracy' on presence detection.
- The HuggingFace model has a known format incompatibility with the sensing server (JSONL vs binary RVF magic bytes) and falls back to null output rather than heuristic mode — meaning the primary advertised ML feature doesn't actually work end-to-end with the published weights.
- The repository is clearly AI-generated or AI-assisted at scale (105-module catalog, 115 ADRs, hundreds of agent/skill/command config files in .claude/) — the scope of documented features vastly outpaces what could plausibly be implemented and tested by a real team, making it hard to know what actually runs.
- Pose estimation (the headline feature) has PCK@20 of 2.5% with no trained keypoint weights shipped yet — the pipeline exists but the data collection phases P7-P9 are pending, so 'see through walls' pose tracking is aspirational marketing, not a working feature today.