VAHA — Offline-First Physical AI Voice Capture Device

Arduino Physical AI Challenge India 2026 — Project Report & Repository

Organised by Robu.in × Arduino

Team Name: VAHA | Team ID: APC-2026-AP-13507 | Track: Smart Homes / Consumer AI
Institution: Dharanova Pvt Ltd, Visakhapatnam, Andhra Pradesh, India
Developer: Ritesh Bonthalakoti (ritesh@dharanova.com) — Solo Developer

   

---

📺 Demo & Product Showcase

📺 Click above or visit https://youtu.be/_Y_hMeHslhI to watch the full physical demonstration.

1. VAHA Hardware Product — by Dharanova Pvt Ltd

2. Companion App — Home Dashboard

---

⏳ Project Development Timeline

---

🌌 Project Overview & Problem Statement

Creative minds don't switch off. Ideas show up at the office, school, the park, a beach, in transit — anywhere you'd normally have a phone or notebook close by. There's one place that isn't true: the bathroom, where the mind keeps wandering with no way to write anything down. VAHA closes that gap — built for the one place your other devices can't follow you, so an idea never has to wait.

Key Innovations:

1. Hands-free Voice Capture: Say "Marvin" to start recording thoughts without touching a phone or unlocking a device.
2. Environmental Context Enrichment: Stamps every voice note with temperature, humidity, TVOC air quality, and water flow rate at that exact moment.
3. 100% On-Device & Offline Privacy: All wake-word detection, speech-to-text, sensor processing, and storage run locally on the Arduino UNO Q. No audio or text ever touches the cloud.

---

🏗️ System Workflow & Architecture

The Arduino UNO Q serves as the central bridge, running an on-device Linux OS environment alongside an MCU microcontroller core to merge local voice AI processing with physical sensing.

graph TD
    subgraph UNO_Q ["Arduino UNO Q Board"]
        Sensors["DHT22 / AGS02MA / Flow"] -->|Raw Readings| MCU["microcontroller Core"]
        MCU -->|sensors_get Bridge API| LinuxOS["Linux OS Runtime"]
        Mic["CS202 USB Microphone"] -->|48 kHz Mono PCM| LinuxOS
    end

    subgraph Linux_Backend ["Linux OS Runtime (Python Backend)"]
        LinuxOS -->|Edge Impulse Model| VAD["Marvin Wake Detect"]
        VAD -->|Active Recording| AudioLoop["Audio Capture State Machine"]
        AudioLoop -->|Stop Keyword / Silence| StopDetect["im_done detection"]
        StopDetect -->|faster-whisper base.en int8| STT["On-Device Transcription"]
        STT -->|Assemble Capture| Storage["Local storage: audio + JSON + checksum"]
        Storage -->|FastAPI Edge Server| Webserver["FastAPI endpoint: 8080"]
        Storage -->|Optional Sync| Notion["Notion Client API"]
        AudioLoop -->|Piper TTS| Speaker["PAM8403 Audio Out"]
    end

    subgraph Mobile_App ["Mobile Companion (Expo / React Native)"]
        Webserver -->|Wi-Fi HTTP Sync| SyncService["Sync Service"]
        SyncService -->|Verify Checksum| SQLite[("Drizzle SQLite DB")]
        SQLite -->|Dynamic Card List| AppUI["React Native Viewports"]
        AppUI -->|Interactive SVG charts| GraphPanel["Telemetry Panel"]
    end

Loading

Step-by-Step Data Flow:

1. Wake Word Detection: The user says "Marvin". The on-device Edge Impulse model (new-marvin.eim) detects it and triggers a capture_started event over WebSockets.
2. Audio Capture: Raw audio is recorded at 48 kHz (16-bit PCM) with noise reduction.
3. Stop Trigger: Recording ends when the stop phrase "im_done" is detected (threshold 0.80) or after 10 seconds of silence.
4. On-Device STT: Audio is downsampled to 16 kHz and transcribed locally using faster-whisper (base.en, int8 quantized, running on CPU).
5. Sensor Sync: Environmental readings are pulled via the sketch's sensors_get() bridge call and packaged into a JSON metadata payload.
6. Local Storage: The capture package (audio.wav, transcript.json, metadata.json, checksum.md5) is written locally to captures/YYYY/MM/DD/uuid/ and optionally synced to Notion.
7. Mobile Sync: The companion React Native app pulls the captures over local Wi-Fi, verifies the MD5 checksums, inserts records into Drizzle SQLite, and issues a purge command to clear the physical device storage.

---

📐 System Block & Circuit Diagrams

Figure 1. VAHA System Block Diagram

Figure 2. VAHA Main Circuit Schematic

---

📷 Complete Project Image Gallery

📱 Mobile Companion Application Views

Home View — Synced Notes

Device View — Live Telemetry

Insights View — Trends

Notes View — Searchable Logs

🛠️ Hardware Build & Sensor Test Rigs

Development Workspace & Assembly Layout	Arduino UNO Q & Microphone Setup
DHT22 & AGS02MA TVOC Sensor Wiring	Water Flow Sensor In-Line Testing
Hardware Components Suite	Enclosed VAHA Device Sensor Port

---

🔌 Hardware BOM & Wiring

Bill of Materials (BOM)

• Microcontroller: Arduino UNO Q (ABX00087) — 4GB RAM / 32GB storage
• Climate Sensor: DHT22 Temperature & Humidity Sensor
• Air Quality Sensor: AGS02MA TVOC Air Quality Sensor (I2C)
• Water Sensor: Hall-effect Pulse Water Flow Sensor (7.5 pulses/L/min)
• Audio Input: USB Lavalier Microphone (auto-detected via CS202 adapter)
• Audio Output: PAM8403 Audio Amplifier + 4Ω 3W Speaker (chime + Piper TTS output)
• Peripherals: Portronics USB-C multiport hub, 2x Mini Breadboards, 10000 mAh Power Bank

Pin Connection Map

Sensor/Module	Module Pin	Arduino UNO Q Pin	Connection Type	Description
DHT22	VCC	5V	Power	Temperature & Humidity Sensor
DHT22	DATA	Pin D2	Digital Input	Climate telemetry signal line
DHT22	GND	GND	Ground	Common ground
Water Flow	VCC	5V	Power	Hall-effect pulse sensor
Water Flow	SIG	Pin D3	Digital Interrupt	RISING edge interrupt pulse counter
Water Flow	GND	GND	Ground	Common ground
AGS02MA	VCC	3.3V	Power	TVOC air quality sensor
AGS02MA	SDA	Pin A4 (SDA)	I2C Data	Communicates at 20kHz clock
AGS02MA	SCL	Pin A5 (SCL)	I2C Clock	-
AGS02MA	GND	GND	Ground	Common ground
PAM8403	5V / GND	5V / GND	Power	Speaker amplifier module
PAM8403	Audio In	Audio Out (Analog)	Analog Input	Voice prompt TTS output from Uno Q
Speaker	L+ / L-	Speaker Outputs	Analog Output	4Ω 3W audio transducer output

---

🤖 AI / ML Model Specifications

Layer / Task	Model Used	Platform / Runtime	Training & Dataset
A: Wake-word Spotting	"Marvin" (new-marvin.eim)	Edge Impulse Runner	Trained on 48 custom voice logs augmented into 1,800 sample iterations (98.7% accuracy).
B: Stop-phrase Spotting	"im_done" (new-marvin.eim)	Edge Impulse Runner	Trained on custom-recorded voice command datasets.
C: Speech-to-Text (STT)	faster-whisper small (277M)	CTranslate2 (int8, CPU)	Fine-tuned English speech model running 100% offline on-device.

---

📊 Verification & Real Bathroom Noise Results

Testing Highlights:

• Verified under real bathroom noise — 34 capture trials: Wake word and stop phrase were exercised across 34 capture attempts in an actual bathroom environment with the exhaust fan running and the tap open. Detection held reliably throughout at 0.75 / 0.80 thresholds.
• End-to-end flow verified: Wake word ➔ Recording ➔ Stop phrase / VAD cutoff ➔ On-device transcription ➔ Local SQLite save ➔ Mobile sync.
• Checksum Verification: Confirmed on every capture package synced to the companion app with zero corrupted transfers.

Performance Metrics:

• Sync-Initiation Latency: ~250 ms average per capture package.
• Transfer Throughput: ~1.5 MB/s over local Wi-Fi.
• CPU Utilization: ~30% peak CPU usage on the Arduino UNO Q during local Whisper inference.
• WebSocket Telemetry Latency: <10 ms latency for real-time sensor updates.
• Database Write Latency: ~5 ms per SQLite transaction in the mobile app.

---

📱 Mobile Companion APK

The compiled Android companion application is available in the repository:

• 📥 Download ARM64 App Build (v1.1.0) (Optimized for arm64-v8a)
• 📥 Download ARMv7 App Build (v1.1.0) (Optimized for armeabi-v7a)

---

⚙️ Setup & Installation

1. Microcontroller Firmware Flash

1. Install Arduino CLI or Arduino IDE.
2. Open sketch/sketch.ino.
3. Install dependencies: DHT library and Arduino_RouterBridge library.
4. Upload the sketch to the Arduino UNO Q board.

2. Python Backend Edge Runtime

1. Install Python 3.12 on the Uno Q Linux workspace.
2. Navigate to python/:

   cd python
   python -m venv venv
   # Activate:
   .\venv\Scripts\activate  # Windows
   source venv/bin/activate # Linux/macOS

3. Install dependencies:

   pip install -r requirements.txt

4. Copy .env.example to .env and fill in API keys (Notion database IDs, Groq token for fallback).
5. Start the edge server:

   python main.py

3. Mobile Companion Application (React Native)

📥 Download Compiled Android APK (Release v1.1.0)

You can directly download and install the standalone release APK binary from GitHub Releases:

• GitHub Release v1.1.0: https://github.com/dharanova-code/vaha/releases/tag/v1.1.0
• Direct APK Download: Download vaha-v1.1.0.apk (88.7 MB)

🛠️ Building / Running Locally

1. Install Node.js (LTS version).
2. Navigate to mobile/:

   cd mobile
   npm install

3. Run the development server:

   npx expo start

---

📄 License & Intellectual Property

Proprietary. All rights reserved. Code licensed under custom terms for Dharanova Private Limited.