AURA — on-device intelligence for microcontrollers

What if the microcontroller itself could think?

AURA turns an ESP32 into a small agent that answers natural prompts — English or Indonesian — entirely on-device. No API keys, no cloud round-trip, no hallucinated facts: AURA answers from curated, swappable knowledge models and declines everything outside its domain.

You:  nyalakan pin 5
AURA: ⚙ GPIO 5 → ON (HIGH)

You:  cek suhu sekarang
AURA: ⚙ current temperature: 36.4 °C (internal chip sensor)

You:  what is a plc
AURA: 📚 PLC — Programmable Logic Controller
      A PLC is a ruggedized industrial computer that controls machines...

You:  what is the meaning of life
AURA: I must decline — my loaded knowledge model "automation v3"
      doesn't cover that topic.

Architecture: two models + a wasm runtime

Layer	What	Swappable?
PRIMARY model	Greetings + hardware integration: GPIO actuators, digital/analog sensors, PWM, temperature, I2C discovery. Enforces a safe-pin allowlist.	No — compiled into firmware, survives every model swap
ADDITIONAL model	A TOON knowledge pack in LittleFS: weighted-keyword retrieval, decline threshold, per-model temperature for varied phrasing.	Yes — web upload, URL fetch, or data/ folder
wasm3 runtime	Runs portable WebAssembly modules on-chip (fib 27, or upload your own .wasm).	Yes — modules are just files

Quick start (PlatformIO)

; platformio.ini
[env:esp32s3]
platform = espressif32
board = lolin_s3_mini
framework = arduino
board_build.filesystem = littlefs
board_build.partitions = huge_app.csv
lib_deps = https://github.com/letjek/AURA.git

#include <AURA.h>

void setup() {
  AuraClass::Config cfg;
  cfg.ssid = "YourWiFi";     // optional — omit for AP-only mode
  cfg.pass = "YourPassword";
  AURA.begin(cfg);
}

void loop() { AURA.loop(); }

Flash, then open http://192.168.4.1/ (join the AURA-Device hotspot, password aura1234) or http://aura.local/ on your LAN. Full working project: examples/ESP32S3-SuperMini.

You can also ask AURA from code: String a = AURA.ask("cek suhu");

Prompt reference

Primary model (always available) hw · pin 5 on|off · pin 5 read · adc 4 · pwm 5 128 · led on|off · temp / cek suhu · i2c scan — plus greetings (halo, hi, apa kabar) and Indonesian verbs (nyalakan, matikan, baca, berapa).

Knowledge model — any other prompt is matched against the loaded pack; below-threshold prompts are declined. model shows what is loaded.

Utility — status, fib <n> (WebAssembly on-chip), echo <text>, help.

Knowledge models (TOON)

name: automation
version: 3
author: Professor Claude
threshold: 2
temperature: 0.7
entries[23]{t,k,a}:
  "PLC — Programmable Logic Controller","plc:2 programmable:1 logic:1","A PLC is..."

• k is keyword:weight ...; a prompt's matched weights are summed and compared to threshold — below it, AURA declines.
• temperature (0–1) varies the presentation (openings, footers, related- topic hints). Facts never vary.
• Install: web page → Install & reboot, paste a raw URL → Fetch & reboot, or drop the file at data/model.toon and pio run -t uploadfs.
• Sample pack: extras/models/automation.toon.

HTTP API

Endpoint	Method	Purpose
/api/prompt	POST (text)	ask AURA
/api/model	GET / POST	download / install knowledge model
/api/model/fetch	POST url=	device downloads a model itself
/api/model/info	GET	one-line model summary
/api/wasm	POST (multipart)	run an uploaded .wasm on-chip

Compatibility

ESP32-class devices with WiFi, ≥4 MB flash and a LittleFS partition. Reference board: ESP32-S3 Super Mini (safe-pin list is S3-based). Small AVR boards (Arduino Nano/Uno) are not supported — not enough RAM/flash for the wasm runtime and web stack.

Ecosystem (roadmap)

• AURA-drivers — public registry + CI that compiles sensor/actuator integrations (AHT10, BMP280, …) into downloadable .wasm driver modules. Firmware exposes i2c/gpio/delay host functions; drivers export init/read. I2C/SPI sensors fit wasm; timing-critical protocols (DHT11 1-wire) stay in the primary model.
• AURA-knowledge — public registry of TOON knowledge packs.
• Public and private entries: private models/drivers encrypted per-creator, unlocked on-device with a key from the creator.
• The original cloud-LLM prototype (multi-LLM, Telegram, MQTT) lives in extras/legacy-prototype and is the base for a future optional cloud-reasoning mode.

License

MIT