Sees without cameras. Hears without microphones. Knows without asking.

---

"The æther is not empty. It is full of perception."

---

Important

Æther is the first RF sensing system built in Gleam/OTP. It turns invisible radio waves — WiFi, BLE, UWB, radar — into human perception: pose, vital signs, presence, activity. No cameras. No wearables. Just physics.

---

⚡ Quick Start

import aether
import aether/core/types.{Zone}
import aether/sensor

let assert Ok(hub) =
  aether.space("home")
  |> aether.add_zone(Zone("sala", "Sala", #(0.0, 0.0, 5.0, 4.0), 0.0, 3.0))
  |> aether.add_sensor(sensor.wifi_csi(
    host: "192.168.1.50", port: 5000,
    antennas: 3, subcarriers: 56, sample_rate: 100,
  ))
  |> aether.with_api(8080)
  |> aether.start()

curl localhost:8080/api/perceptions

{
  "perceptions": [
    {"type": "pose", "keypoints": [...], "skeleton": "coco17", "confidence": 0.92},
    {"type": "vitals", "heart_bpm": 72.3, "breath_bpm": 16.1, "hrv": 45.2},
    {"type": "presence", "total_occupants": 2},
    {"type": "activity", "label": "walking", "confidence": 0.85}
  ]
}

🏗️ Architecture

%%{init: {'theme': 'base', 'themeVariables': { 'primaryColor': '#e94560', 'primaryTextColor': '#fff', 'primaryBorderColor': '#0f3460', 'lineColor': '#0f3460'}}}%%
flowchart LR
    subgraph Sensors["📡 RF Sensors"]
        ESP[ESP32-S3<br/>WiFi CSI]
        RADAR[mmWave<br/>Radar]
        BLE[BLE<br/>Beacons]
        UWB[UWB<br/>DW3000]
    end

    subgraph Aether["🧠 Æther (Gleam/OTP)"]
        direction TB
        SA[Sensor Actors<br/>UDP + parse]
        PIPE[Conditioner<br/>Rust NIFs]
        ORCH[Orchestrator<br/>Ring Buffer + AveCSI]
        BRAIN[Brain NIF<br/>Foundation Model]
        TRACK[Kalman Tracker<br/>Multi-Person]
    end

    subgraph Output["💡 Perception"]
        API[REST API]
        WS[WebSocket<br/>Real-time]
        EVENTS[Events<br/>Fall/Enter/Leave]
    end

    ESP & RADAR & BLE & UWB -->|UDP| SA
    SA --> PIPE --> ORCH --> BRAIN --> TRACK
    TRACK --> API & WS & EVENTS

Loading

Property	Value
Language	Gleam + Rust NIFs + Erlang FFI
Runtime	BEAM/OTP 28
Tests	72 passing
Latency	< 15ms end-to-end
Lines	5,800+ (Gleam, Rust, Erlang, C)

---

🔬 Signal Processing (Rust NIFs)

Algorithms validated in 2025-2026 research papers:

NIF	Algorithm	Paper	Scheduler
tsfr_calibrate	Phase unwrap + detrend + Savitzky-Golay	TSFR (2023)	Normal
hampel_filter	Outlier detection via MAD	Standard DSP	Normal
butterworth_bandpass	Zero-phase IIR cascaded biquads	Standard DSP	Normal
savgol_filter	Polynomial least-squares smoothing	Savitzky-Golay (1964)	Normal
avecsi_stabilize	Sliding window frame averaging	CSIPose (IEEE TMC 2025)	Normal
spotfi_aoa	MUSIC-based AoA estimation	SpotFi (SIGCOMM 2015)	Normal
foundation_infer	Multi-task foundation model	AM-FM / X-Fi (2026)	DirtyCpu
cross_modal_fuse	Attention-weighted sensor fusion	X-Fi (2025)	DirtyCpu

Signal NIFs run on normal BEAM schedulers (< 1ms). Brain NIFs run on dirty CPU schedulers to avoid blocking.

---

🧠 Foundation Model

One model. Five decoder heads. Any RF modality.

%%{init: {'theme': 'base', 'themeVariables': { 'primaryColor': '#e94560', 'primaryTextColor': '#fff', 'primaryBorderColor': '#0f3460', 'lineColor': '#0f3460'}}}%%
flowchart TB
    EMB[Fused Embedding<br/>Cross-Modal Attention] --> POSE[🦴 Pose<br/>17 COCO Keypoints]
    EMB --> VITALS[❤️ Vitals<br/>Heart + Breath + HRV]
    EMB --> PRESENCE[👥 Presence<br/>Occupancy + Zones]
    EMB --> ACTIVITY[🏃 Activity<br/>idle/walk/fall/wave]
    EMB --> LOCATION[📍 Location<br/>3D + Accuracy]

Loading

Decoder	Output	Inspired by
Pose	17 keypoints + velocity	GraphPose-Fi, VST-Pose (2025)
Vitals	Heart BPM, breath BPM, HRV	PulseFi (2025), RoSe (2026)
Presence	Zone occupancy count	AM-FM (2026)
Activity	Label + confidence	X-Fi (2025)
Location	3D position + accuracy (m)	Geometry-Aware (2026)

---

🎯 Multi-Person Tracking

• Kalman filter per person per axis (x, y, z) — smooths noisy RF measurements
• Nearest-neighbor association — links detections to tracked persons (< 2m threshold)
• Dead reckoning — predicts position between frames using estimated velocity
• Stale removal — drops persons not seen for configurable timeout
• Zone events — PersonEntered, PersonLeft, FallDetected, VitalsAlert

---

🌐 API

REST

GET  /api/health        → {"status": "ok", "version": "0.1.0", "name": "aether"}
GET  /api/perceptions   → {"perceptions": [...], "count": 5}
GET  /api/sensors       → {"sensors": [...]}

WebSocket

ws://localhost:8080/ws/stream → real-time perception push (JSON)

Supports ping → pong heartbeat. Perceptions are pushed automatically via OTP subject subscription.

---

📡 Supported Sensors

Sensor	Transport	Signal Type	Capabilities
ESP32-S3	UDP	WiFi CSI (56 subcarriers)	Pose, vitals, presence, through-wall
IWR6843	UDP	mmWave range-Doppler	Micro-movements, vitals, gesture
DW3000	UDP	UWB CIR	Centimeter-level localization
BLE beacons	UDP	RSSI	Coarse room-level presence
Any WiFi	UDP	RSSI (degraded)	Basic presence detection
Custom	UDP/Serial/TCP	User-defined	Extensible via UserDefinedSignal

🔧 ESP32 Firmware

Included in firmware/esp32-csi-node/. Captures WiFi CSI (LLTF + HTLTF + STBC) and streams via UDP.

cd firmware/esp32-csi-node
idf.py menuconfig  # Set WiFi SSID, password, hub IP, port
idf.py build
idf.py -p /dev/ttyACM0 flash monitor

Configurable via Kconfig: SSID, password, hub IP, hub port, sample rate.

---

📚 Research Foundation

Built on state-of-the-art 2025-2026 papers:

Paper	Year	Key Contribution
AM-FM	2026	Foundation Model for Ambient Intelligence Through WiFi
X-Fi	2025	Cross-modal transformer — 24.8% MPJPE reduction
WiFlow	2026	Lightweight axial attention + TCN for pose
CroSSL	2026	Station-wise masking for sensor robustness
GraphPose-Fi	2025	Per-antenna GCN + self-attention decoder
VST-Pose	2025	Velocity-Integrated Spatio-Temporal Attention
CSIPose	2025	AveCSI stabilization for through-wall sensing
PulseFi	2025	ESP32 CSI heart rate pipeline
LatentCSI	2025	CSI → Stable Diffusion latent space
Diffusion²	2026	3D environments → RF heatmaps

---

🏛️ OTP Supervision

Space (Supervisor) ── OneForOne
├── Orchestrator (worker)
│   ├── Ring Buffer (AveCSI stabilization)
│   └── Brain NIF (foundation model)
├── SensorSupervisor
│   ├── Sensor("esp32-sala") + UDP listener
│   ├── Sensor("esp32-quarto") + UDP listener
│   └── Sensor("radar-entrada") + UDP listener
├── PerceptionAggregator
│   ├── Kalman Tracker (per-person state)
│   └── Zone Manager (transition events)
└── ApiServer (Mist HTTP/WS)

Sensor crashes → only that sensor restarts. Orchestrator crashes → restarts fresh, sensors auto-reconnect via bridge. Brain NIF crashes → dirty scheduler isolates, supervisor restarts.

---

📂 Project Structure

aether/
├── src/aether.gleam                  # Public API — space() |> add_sensor() |> start()
├── src/aether/
│   ├── core/                         # Types (Vec3, Zone, Signal), errors, math
│   ├── sensor/                       # OTP actors, UDP listener, CSI frame parser
│   ├── condition/                    # Pipeline (NIF-wired), ring buffer, AveCSI
│   ├── fusion/                       # Temporal sync, cross-modal fusion engine
│   ├── orchestrator.gleam            # Heart — signal → condition → brain → perception
│   ├── track/                        # Kalman filter, aggregator, zone events
│   ├── serve/                        # REST API, WebSocket streaming, JSON codec
│   └── nif/                          # Gleam FFI bindings to Rust NIFs
├── native/
│   ├── aether_signal/                # Rust NIF: TSFR, Hampel, Butterworth, SpotFi
│   └── aether_brain/                 # Rust NIF: foundation model (5 decoders)
├── test/                             # 72 tests across all layers
└── firmware/esp32-csi-node/          # ESP32-S3 CSI capture + UDP streaming

---

🛠️ Build

📋 Prerequisites

Tool	Version
Gleam	>= 1.14
Erlang/OTP	>= 28
Rust	>= 1.94
ESP-IDF	>= 5.x (optional, for firmware)

# Clone
git clone https://github.com/gabrielmaialva33/aether.git && cd aether

# Build Gleam
gleam build

# Build Rust NIFs
make nif

# Test everything
gleam test
# 72 passed, no failures

# Run
gleam run

---

📊 Stats

Metric	Value
Gleam modules	21
Rust modules	11
Erlang FFI modules	6
Total lines	5,800+
Tests	72
Rust tests	14
API endpoints	3 + WebSocket
Signal NIFs	6
ML decoder heads	5

---

📜 License

MIT

---

Built with invisible waves and visible ambition.