Skip to main content
Camper van interior at night with the Peregrine voice assistant on a shelf, aurora visible through the window

Peregrine

Say "hey Peregrine" and turn the lights off. Ask what the battery is at. Ask how full the fresh water tank is, or whether the trailer is level. All of it runs on a chip inside the van. No cloud, no account, no mic feed leaving the rig.

What's Inside  /  Module

The Voice Assistant

Peregrine is a fully offline voice assistant for your rig. A custom wake word, local speech-to-text, a language model running on a dedicated NPU, and local text-to-speech. MQTT plugs it straight into every other module on the bus so your voice becomes another input.

Built around the Radxa Dragon Q6A, a single board computer with a Qualcomm QCS6490 SoC and a Hexagon NPU. The LLM runs on the NPU at about twelve tokens per second. The wake word, STT, and TTS run on the CPU. Add a small buck converter to hold the 12 V rail steady so a fully charged lithium pack can't push the board past spec, a Deutsch DTM4 for power, and a two-piece 3D printed case. Done.

Radxa Q6A Hexagon NPU Llama 3.2 openWakeWord Piper TTS MQTT
View on GitHub
TrailCurrent Peregrine enclosure rendering

At a Glance

Everything the assistant needs, on a single board inside a small black case.

Compute

Radxa Dragon Q6A. Qualcomm QCS6490 with eight Kryo cores, 8 GB LPDDR5, and a Hexagon NPU that runs the LLM at roughly twelve tokens per second. Radxa OS Noble 24.04 underneath, no desktop, no display server, just the voice pipeline.

Voice Pipeline

A custom hey_peregrine wake word from openWakeWord, faster-whisper base.en in INT8 for speech-to-text, Llama 3.2 1B on the Hexagon NPU for reasoning, and Piper TTS (libritts_r-medium) for voice output. Everything on-device.

Offline by Default

No cloud services. No account. The mic feed never leaves the vehicle. If the cell bars are gone you still get lights, temperature, battery, and general questions answered. MQTT over TLS handles device control through the TrailCurrent bus.

Bill of Materials

Everything you need to build one Peregrine, top to bottom.

Qty Part Description Source
1 Radxa Dragon Q6A Qualcomm QCS6490 SBC. Eight Kryo cores, 8 GB LPDDR5, Hexagon NPU for on-device LLM inference, dual gigabit ethernet, four USB 3.0 ports, and a dedicated 3-pin 12 V power header separate from the main GPIO pins. Radxa OS Noble 24.04. Radxa
1 M.2 2230 NVMe SSD Holds Radxa OS Noble 24.04 (based on Ubuntu 24.04), the Peregrine Python venv, the NPU LLM model (Llama 3.2 1B for Hexagon), the Piper TTS voice, and the custom wake word model. The golden image is about 6 GB, so 128 GB is plenty of headroom. Must be the 2230 form factor: the Dragon Q6A only has a single M.2 M Key slot and it's sized for 2230. The more common 2280 drives will not fit. Any brand
1 LM2597 Buck Converter Module Wide-input step-down module wired as a 12 V regulator. The Radxa Dragon Q6A takes 12 V on its power header, but a fully charged 4S lithium pack sits at 14.6 V and an alternator can spike higher. The buck clamps whatever the vehicle bus is doing down to a clean 12 V so the board stays in spec. A couple of electrolytics on the output smooth the rail and a trim pot sets the 12 V set point. Generic
1 Jabra Speak USB Speakerphone One USB device handles both the microphone array and the speaker, with on-device echo cancellation and noise suppression that make a huge difference in a noisy cabin. Plugs into a single USB-A port on the Radxa. This is the only combination we've tested end to end on Peregrine. Audio on Linux is a rabbit hole and getting a reliable ALSA path out of a random USB mic and a random powered speaker is not worth the fight, so we standardized on the Jabra and that's what the setup script expects. Jabra
1 Deutsch DTM04-4P Connector Sealed 4-pin receptacle on the port side of the case for 12 V power and (optionally) the TrailCurrent CAN pair. Peregrine itself talks MQTT over ethernet or WiFi; the CAN pins are unused on this module. TE Connectivity
1 Case Bottom 3D printed. 95 × 121 × 23.5 mm (with mounting tabs). Integrated standoffs for the Radxa Dragon Q6A board and the buck converter, plus an opening for the DTM4 connector and a port cutout that frames the Radxa's four USB-A ports and the RJ45 ethernet jack. ABS or ASA recommended. 3D printed (STL below)
1 Case Cover 3D printed. 95 × 100 × 20.1 mm. Embossed TrailCurrent logo on the top face and four clearance holes for the cover screws. Print this one with the logo face up for the cleanest finish. ABS or ASA recommended. 3D printed (STL below)
6 M2.5 × 3 mm Machine Screws (internal mount) Four secure the Radxa board to its standoffs, two secure the LM2597 buck converter to its standoffs. Pan-head, stainless or zinc-plated. Hardware store
4 M2.5 × 6 mm Machine Screws (cover) Secure the cover to the case bottom at the four corners. Pan-head, stainless or zinc-plated. Hardware store

Technical Drawings

Orthographic views of the full assembly, straight out of the FreeCAD model.

Peregrine top view with dimensions

Top View

Embossed TrailCurrent logo on the cover, four cover screws at the corners, and four mounting tabs on the long ends. The DTM4 connector pokes out on the port side.

Peregrine front view with dimensions

Front View

Vent slots run the full length of one long side so the Radxa SoC can breathe. No fan: passive cooling is enough for the idle-most-of-the-time voice assistant workload.

Peregrine port side with dimensions

Port Side

The business end. Deutsch DTM 4-pin power connector, four USB-A ports for the Jabra Speak and accessories, and a gigabit ethernet jack that uplinks Peregrine to the rest of the network.

Peregrine blank side with dimensions

Blank Side

The face that usually sits against a bulkhead. No ports, no vents. Keeps the install clean and the label-free side toward the cabin.

Overall dimensions: 98 × 121 × 27 mm. Source CAD lives in the CAD folder of the Peregrine repo.

3D Printed Parts

Two pieces, ABS, set up as a multi-filament print so the TrailCurrent logo on the cover comes out in brand green. A ready-to-print project lives on Makerworld, and the raw .3mf and STL files are in the CAD folder of the repo.

Peregrine case bottom, 3D printed

Case Bottom

The main body. Integrated standoffs hold the Radxa Dragon Q6A and the LM2597 buck converter, a long vent array runs the full length of one side to shed SoC heat, and the short end frames the DTM4 power connector and the Radxa's port cluster. Four mounting tabs at the corners for the install. Prints with the open top facing up, so the tub sits on the build plate and the walls grow straight up.

  • Dimensions: 95 × 121 × 23.5 mm
  • Material: ABS (or ASA)
  • Nozzle: 0.4 mm
  • Layer height: 0.2 mm
  • Walls: 6 perimeters
  • Infill: 100%
  • Orientation: Open side up (standoffs point at the nozzle)
  • Supports: Tree (auto), enabled
  • Nozzle temp: 270 °C (260 °C first layer)
  • Bed temp: 90 °C
  • File: BodyTrailCurrentPeregrineCaseBottomV10.stl
Makerworld STL / .3mf on GitHub
Peregrine case cover, 3D printed

Case Cover

The lid. Embossed TrailCurrent logo on the top face and four clearance holes for the cover screws. Set up as a two-filament print: body in the case color, logo swap to TrailCurrent green on the embossed face. Prints on its edge, not flat. Standing the cover up means the multi-color layers run vertically across the logo face, which gives a much cleaner color boundary than trying to print the logo on a horizontal top surface. The profile in the .3mf handles the orientation and the filament mapping for you.

  • Dimensions: 95 × 100 × 20.1 mm
  • Material: ABS (or ASA), two filaments (case color + green for logo)
  • Nozzle: 0.4 mm
  • Layer height: 0.2 mm
  • Walls: 6 perimeters
  • Infill: 100%
  • Orientation: On edge, logo face perpendicular to the build plate
  • Supports: Tree (auto), enabled
  • Nozzle temp: 270 °C (260 °C first layer)
  • Bed temp: 90 °C
  • Files: BodyTrailCurrentPeregrineCaseTopNoLogoV2.stl + PeregrineCaseTopV2.stl (logo insert)
Makerworld STL / .3mf on GitHub

Assembly

Board-mount screws, drop in the Radxa and the buck, wire the DTM4, close the cover.

  1. 1

    Print the case

    Easiest path: open TrailCurrentPeregrine.3mf from the CAD folder in Bambu Studio (or Orca) and print both plates. ABS, 0.2 mm layers, 100% infill, 6 walls, tree supports on the bottom. If you're slicing from STL, match those settings; the cover prints flat with no supports.

  2. 2

    Set up the Radxa board

    On the bench, flash Radxa OS Noble 24.04 onto the Dragon Q6A and boot it. Clone the Peregrine repo, run setup/setup-board.sh as root, and let it install the NPU runtime, create the assistant user, download the Llama 3.2 NPU model and the Piper voice, install the wake word, and register the systemd services. It's idempotent, safe to re-run.

  3. 3

    Mount the Radxa in the case

    Lower the Radxa Dragon Q6A into the bottom of the case so its port cluster lines up with the port-side cutout. Drive four M2.5 × 3 mm screws through the board's mounting holes and into the case's integrated standoffs. Snug, not cranked down; the plastic threads are gentle.

  4. 4

    Drop in the buck converter

    Set the LM2597 buck converter into its pocket next to the Radxa and secure with two more M2.5 × 3 mm screws. Before wiring it up, trim the pot on the buck to exactly 12.0 V with a meter. A fully charged 4S lithium pack can sit at 14.6 V and the Radxa's 12 V input is unforgiving about anything much above spec.

  5. 5

    Wire the DTM4 to the Radxa power pins

    Press the Deutsch DTM04-4P receptacle into its opening and land vehicle 12 V / ground on the input side of the buck. Run the regulated 12 V output from the buck to the Radxa Dragon Q6A's three-pin power header. This header is a separate block of three pins sitting right next to where the buck lands in the case, not part of the main GPIO header, and it took us a while to track it down in the Dragon Q6A docs. Route cables so they clear the cover standoffs.

  6. 6

    Close the case

    Drop the cover on. Secure with four M2.5 × 6 mm screws at the corners. Snug, not cranked down; the plastic threads are gentle.

  7. 7

    Plug in the Jabra Speak

    Connect the Jabra Speak USB speakerphone to one of the Radxa's USB-A ports. One cable gives you both the mic array and the speaker, with echo cancellation already handled in the Jabra itself. Confirm it shows up with arecord -l and aplay -l. This is the only audio hardware we've validated on Peregrine, and the setup script expects it.

  8. 8

    Say hey Peregrine

    Power it up. Watch the boot over ethernet with journalctl -u voice-assistant -f. Once the wake word model loads, say "hey Peregrine, what's the battery at" and you should hear Piper answer back. Set the MQTT broker in ~/assistant.env to give it access to lights, relays, and sensor data through the TrailCurrent bus.

Voice Pipeline

Four local models, one simple loop. No network calls, no cloud round-trip.

1. Wake Word

A custom hey_peregrine openWakeWord model trained on a mix of real recorded positives and TTS-generated variants. Runs continuously on a single CPU core, sub-one-percent load, ~200 MB of audio buffer. Tuneable threshold, default 0.85. Negative clips of ambient cabin noise keep false positives near zero.

2. Speech-to-Text

faster-whisper base.en in INT8, running on the CPU. Records two to three seconds of audio after the wake word fires and transcribes locally in well under half a second. No audio ever leaves the box. No audio ever hits the filesystem unless you explicitly turn on debugging.

3. Intent and LLM

Fast regex-based intent matching handles the common commands directly: lights on / off, brightness, relays, temperature, battery, location. Anything that doesn't match falls through to Llama 3.2 1B running on the Hexagon NPU over the genie-t2t-run HTTP server. Roughly twelve tokens per second, no round-trip to a data center.

4. Text-to-Speech

Piper TTS (en_US-libritts_r-medium). Synthesizes the response on the CPU in real time and plays it straight out through ALSA. No cloud API, no billing, no latency from a network hop. The whole "listen, understand, answer" loop usually finishes in well under two seconds.

Device Control (MQTT)

Matched intents fire MQTT messages onto the TrailCurrent bus: devices/{id}/set for lights, relays, and thermostat. Torrent (PWM lights) and Switchback (relays) both respond. Brightness only applies to Torrent, relays are on/off. Device names come from the TrailCurrent web UI so "turn off the awning light" just works.

Hardening

The setup script disables the desktop, GPU stack, HDMI, snapd, and any service not strictly needed for the voice loop. CPU governor pinned to performance. Kernel swappiness and dirty-page tuning cut audio jitter. assistant user runs the service with a locked password; there's no interactive login.

Voice Commands

A taste of what the assistant responds to. Unrecognized phrases fall through to the LLM.

Category Example Phrases
Lights on / off "Turn on the lights", "Lights off", "Turn off light 3"
Light brightness "Set lights to 50 percent", "Dim light 2 to 25 percent"
Named devices "Turn on the water pump", "Turn off the awning light"
Everything "Turn off everything", "Turn on all the lights"
Temperature / humidity "What's the temperature?", "How hot is it?", "What's the humidity?"
Battery / energy "What's the battery level?", "How much solar?"
Location "Where are we?", "What's our location?"
Anything else Answered by the on-device LLM. No network, no cloud, no wait.

Build Your Own Peregrine

Every file is in the repository: CAD, wake word training pipeline, Python assistant, NPU LLM server, and the systemd service files. Fork it, build it, make it yours.

Peregrine on GitHub