Sentry Node

The MAX78000 has a hardware CNN accelerator alongside a conventional ARM Cortex-M4 core, meaning it can run convolutional networks at orders of magnitude lower power than a general-purpose CPU. Models can be retrained for the deployment context — detecting people in a yard, livestock in a field, vehicles on a driveway, or specific machinery states on a factory floor. Because inference happens locally, no video or audio ever leaves the device unless you explicitly request it. A CP2102 USB-to-UART bridge is wired in for flashing new models and reading serial logs over a standard USB cable.

The ESP32-S3 (WROOM-1U-N8R8) provides 8 MB of flash and 8 MB of PSRAM, dual-core 240 MHz Xtensa LX7, and an external antenna connector for extended LoRa and Wi-Fi range. It owns the directional audio DSP pipeline, drives the speaker driver and addressable status LEDs, monitors the battery and solar input, and handles all packet framing and routing for the LoRa radio. The split-MCU architecture means inference and systems work never block each other — the AI can stay locked on a target while the radio is busy uploading the previous event.

The HM0360 is a Himax computer-vision sensor optimized for the kind of low-resolution, high-frame-rate detection workloads the MAX78000 thrives on. Its always-on motion-detection mode allows the AI MCU to sleep until something interesting moves into frame, then wake instantly to classify what it is. When a frame matters — a confirmed intrusion, an asset removal event, a livestock tally — the full image can be cached locally and pulled later by an operator over the mesh.

The floodlight has a measurable deterrent effect on its own — the abrupt activation of bright cool-white light when an intruder approaches communicates very clearly that something is watching. For more discreet deployments, the floodlight can be disabled in firmware and replaced with infrared illuminators if covert night vision is required. The LM3410 boost driver runs the LEDs efficiently from a single Li-Po cell, and brightness can be PWM-controlled per event.

Directional audio is what lets the Sentry Node react to events outside the camera's field of view. A footstep, a vehicle door, or an animal vocalization can be localized and either logged as a standalone event or used to cue the visual pipeline — and on a pan-capable mount, to physically rotate the camera toward the sound. Audio classification models can be loaded onto the AI MCU to recognize specific sound signatures: gunshots, glass breaking, machinery faults, livestock distress calls. All audio processing happens locally; recordings are not transmitted unless requested.

Radar fills the gaps a camera cannot. It works in conditions that defeat visible-light sensors — heavy weather, deep night, blowing dust — and it returns physically grounded measurements (Doppler velocity, time-of-flight range) that don't depend on object recognition to be meaningful. The Sentry Node's tracking layer can switch the radar's lock between targets, allowing it to follow a vehicle approaching a gate, then snap to a person stepping out of it. Radar data is fused with camera and audio at the AI layer to produce a single confidence-rated track per detected target.

Dual-band support means the Sentry Node can fall back from a saturated or interfered 915 MHz channel to 2.4 GHz, or run private 2.4 GHz traffic in parallel with public mesh participation on sub-GHz. The radio is sourced as a Waveshare CoreLR1121-HF module with castellated pads for clean PCB integration. Because event payloads are small (timestamp, target class, bearing, range, confidence), a single Sentry Node can broadcast hundreds of detections a day on solar power without ever overrunning its energy budget.

The speaker turns each Sentry Node into a remote loudspeaker on the TrueMesh network. Speak into a Pocket Beacon and your voice is encoded, relayed over LoRa, and played back live through every Sentry Node addressed in the broadcast — effectively a region-wide public address system with no wiring, no amplifier rack, and no cell coverage required. Distribute nodes across a farm, ranch, worksite, or campus and you can call workers in from the far paddock, issue evacuation instructions, or push announcements across the whole property from a device in your pocket. Broadcasts can be targeted to a single node, a group, or the entire mesh, and pre-recorded deterrent messages can fire automatically on intrusion ("you are being recorded, leave the area"). Volume is configurable, and the NS4150B driver supports any 4 Ω speaker swap if a louder field horn is needed.

The MPPT charger automatically tracks the maximum-power operating point of whatever panel is connected, so a 5 W or 15 W panel both run at peak efficiency without any tuning. Two 18650 cells provide enough buffer to ride out a multi-day overcast stretch in nominal operation, and SMD-mounted holders make field cell swaps trivial. For installations with mains power available, a USB-C input lets the same hardware run as a permanent fixture without the solar/battery stack populated. Designs without sun exposure, like factory installations, simply use USB-C.