Take your voice anywhere, transcribe on YOUR hardware.

Tamesh Sivaguru — Sun, 15 Feb 2026 19:44:35 +0000

*This is a submission for the GitHub Copilot CLI Challenge*

🎤 Whisper-Typing Mobile

Privacy-First Speech-to-Text, Anywhere.

I transformed an existing open-source Windows desktop app into a full-scale, cross-platform mobile ecosystem in a single 3-hour session using the GitHub Copilot CLI.

The Challenge: How do you use high-end speech-to-text on a phone while keeping audio data 100% private and avoiding expensive cloud API fees?

The Solution: A self-hosted mobile architecture that leverages your home PC’s GPU power over a secure mesh network.

🚀 The Build at a Glance

Metric	Result
Time to Build	~3 Hours
Lines of Code	~6,500 production lines
Files Created	50+ files
Architecture	8 Phases (Backend ➔ Docker ➔ Mobile ➔ Docs)
Status	Production-Ready

🛠️ The Tech Stack

Frontend: Flutter (Material Design 3) + gRPC Client
Backend: Python 3.13 + FastAPI + gRPC + Protocol Buffers
Inference: faster-whisper + Ollama (NVIDIA CUDA 12.4)
Networking: Tailscale Mesh Network (Encrypted Tunnel)
DevOps: Docker with GPU Passthrough

🏗️ Architecture Overview

┌─────────────────┐
│  Android Phone  │  Push-to-talk recording
│  Flutter App    │  Real-time transcription
└────────┬────────┘
         │ 
         │ gRPC over Tailscale (E2E Encrypted)
         ▼
┌──────────────────────────────────────────┐
│            Docker Container              │
│  ┌──────────────────┐                    │
│  │   gRPC Server    │ Port 50051         │
│  │ (Transcription)  │                    │
│  └──────────────────┘                    │
│  ┌──────────────────┐                    │
│  │  Web Admin Panel │ Port 8080          │
│  │ (Configuration)  │                    │
│  └──────────────────┘                    │
│  ┌──────────────────┐                    │
│  │    Whisper AI    │ Utilizes Home GPU  │
│  │  faster-whisper  │ via NVIDIA CUDA    │
│  └──────────────────┘                    │
└──────────────────────────────────────────┘

✨ Key Features

📱 Mobile App

Push-to-Talk: Simple, intuitive recording interface.
AI Improvement: Integrated Gemini support to polish transcriptions.
Onboarding Wizard: A 4-page setup flow for permissions and connection testing.
History & Clipboard: Session-based history with one-tap copy.

🔌 Backend & Security

Privacy-First: Your voice never touches the cloud. Phone ➔ Tailscale ➔ Your PC.
Hardware Ownership: Use your own NVIDIA GPU for blazing-fast local transcription.
Web Admin: Browser-based monitoring and configuration—no SSH required.
One-Command Deploy: docker-compose up -d and you're live.

🧠 My Experience with GitHub Copilot CLI

This wasn't just "autocomplete"—it was a senior pair programmer. Here is how the CLI changed the game:

1. From Idea to Production in 180 Minutes

Starting from a Windows-only desktop app, I asked the CLI to plan a cross-platform expansion. It designed an 8-phase architecture and helped me execute every single one. Without it, this would have been 2–3 weeks of research and prototyping.

2. Context-Aware Engineering

The CLI didn't just write code; it wrote my code.

It respected my strict linting rules (ruff with ALL enabled).
It matched my Google-format docstring style.
It understood Python 3.10+ type hint requirements automatically.

Example: Copilot knew to use lazy logging to comply with ruff G004 and used Python 3.10+ generics without being prompted.

# Generated by Copilot CLI to match my project standards
from typing import Iterator
import logging

logger = logging.getLogger(__name__)

def transcribe(self, audio: bytes) -> str:
    """Transcribes audio using faster-whisper.

    Args:
        audio: Raw audio bytes in WAV format.
    Returns:
        Transcribed text string.
    """
    logger.info("Processing audio: %s", audio_id) # Validated for lazy logging

3. Documentation as a First-Class Citizen

Normally, documentation is the last thing developers do. The CLI made it part of the flow, generating 7 comprehensive guides (Docker, Backend, User Guides, and QA procedures) that were accurate to the code we just wrote.

💡 The "Aha!" Moments

Parallel Tool Calling: Watching the CLI read three files simultaneously to understand a cross-service bug was eye-opening.
Context Retention: It remembered a Tailscale IP discussion from Phase 1 while we were working on Phase 8.
Error Recovery: When a command failed, it didn't quit; it analyzed the stack trace, proposed a fix, and kept moving.

🔗 Links & Resources

Experimental Code: GitHub Repository (Mobile Branch)
License: MIT

Final Verdict: The Copilot CLI doesn't replace developer judgment—it amplifies it. It handled the mechanical boilerplate with zero fatigue, allowing me to focus entirely on the privacy architecture and user experience.

Forem: Tamesh Sivaguru