Chat with physical space.
Walls talk. Commodity ESP32-S3 nodes. Zero cameras. Jamming-resistant by design. LatentField is an architectural exploration of turning 802.11 CSI into a low-bandwidth event stream any LLM can reason about.
The OLED is the dongle's diary.
Each LatentField node carries a tiny OLED that shows its current state in pixel-art frames. Six of them, cycling. What you see in the hero loop is a real node on a workbench — no renders, no mock-ups. The frames below are stills of the same display.
Six of the eight latents the camera caught in one loop. Energy-state and alert-level cycle less often and didn't make this take.
Three lines, no marketing.
WiFi CSI on commodity hardware
Channel state information from off-the-shelf ESP32-S3 boards. No SDR, no proprietary radio, no licensing. The hardware costs less than a coffee.
Latent-state compression
Continuous CSI is collapsed into eight floats per frame. The world model is the interface; the structured event stream is the API.
Conversational query layer
"Who's in zone 3 right now?" "Did anyone enter the back room after 22:00?" The LLM operates against deterministic events, not raw signal.
How walls become a sensor.
802.11 OFDM splits each channel into 56 CSI subcarriers. Every subcarrier is a complex number — amplitude and phase. Together, they form a CSI matrix that captures the precise way radio waves bend, scatter, and absorb in your room.
Bodies, motion, breathing, occupancy, even slow material changes all leave fingerprints in that matrix. We don't need new hardware to see them — we just need to read what the standard already exposes.
From the CSI matrix, an autoencoder compresses each frame into eight latent fields:
- Presence — is anyone in this zone right now?
- Motion — magnitude, direction, persistence
- Breathing — sub-Hz amplitude modulation
- Occupancy — multi-zone headcount
- Zone activity — what zone, what is happening
- Anomaly — does this match the learned baseline?
- Energy state — passive HVAC/lighting indicators
- Alert level — aggregate event severity
No cameras. No biometric capture. The signal is already there — we just read it.
Four lines of bill of materials.
For comparison: a single UniFi G5 Bullet camera draws ~4W continuous, streams 4 Mbps, and produces 172.8 GB of video per day. A LatentField room produces 128 KB per day, structured.
Inference at the right altitude.
Real-time reflex in the dongle. Multi-room context in the gateway. Historical pattern mining and natural-language queries in the cloud. Bandwidth shrinks at every step.
TinyML on the ESP32-S3
Quantised models run on-device, sub-100ms inference. Output: the eight latent fields, framed as events. Memory: <512 KB flash. CPU: idle most of the time.
Edge aggregation + RuVector
Multi-room state. RuVector embedding store for CSI/event vectors. Dynamic model flashing: push a new TinyML model OTA when a vertical changes (warehouse → laundry → telco).
DarkNOC pattern mining
Historical pattern mining, model training, conversational query layer, fleet management. Queries happen here. The data stays small enough to fit the prompt.
Jam us. Watch us get sharper.
A camera under jamming goes blind. A LatentField sensor under jamming gets better.
Under WiFi jamming, ESP32 nodes retry packet transmissions more aggressively. That means more emissions in the air. More emissions means more CSI samples per second — which means sharper, faster, more confident inference.
When WiFi is fully blocked, the nodes still emit at the physical layer. A rotary display can read those emissions directly, walk the data down a USB serial cable to a laptop, and the floor plan keeps updating in real time. We've tested this. It works.
A jammer cannot touch copper.
USB serial fallback implemented and demonstrated in May 2026.
Failure modes & what we do
- WiFi degraded: CSI quality improves (more retries → more samples)
- WiFi fully jammed: physical-layer read over USB serial → laptop
- Power loss: 200+ hour battery operation in CameraPlus tier
- WAN cut: LoRa SX1262 sub-GHz fallback to a different band
- Everything down: events queue locally; sync on reconnect
We promise a second independent delivery channel. We do not promise guaranteed multi-hop mesh routing.
We're not a camera replacement.
Honesty buys trust faster than spec sheets.
What LatentField doesn't do
- Forensic-grade video evidence
- Identity-grade face or vehicle recognition
- Sub-centimetre localisation
- Compliance recording in regulated jurisdictions where video is the legal artifact
What LatentField does better than cameras
- Continuous presence and flow inference
- Works through walls and in total darkness
- Anomaly detection from ambient state
- Privacy-preserving by construction — no biometric capture
- Power-budget compatible with battery and solar
- Resilient to jamming and network outages
Same physics, three product shapes.
CameraPlus →
RF-first + camera-on-trigger. 6 industrial verticals explored. 91.5% energy reduction observed. The "When Space Talks Back" paper.
Managed monitoringMonitorPlus →
Architectural exploration of monitoring-as-a-service. Four deployment configurations: Edge, Edge + Vision, Resilient Edge, Partner.
Deep architectureTechnology →
ESP-NOW, TinyML, dynamic model flashing, RuVector, the 3-level pipe in detail.