nanosensei

NanoSensei – On-Device AI Skill Coach

Comment

🚀 NanoSensei — Your Personal On-Device Skill Coach Powered by AI

🎯 Inspiration

Learning new skills is still painfully inefficient. Whether you're practicing speech, coding, design, fitness form, or creative skills, most learning apps rely on cloud AI, meaning:

  • High latency
  • Privacy concerns
  • Expensive inference
  • Poor offline support

But ARM-powered mobile devices today are capable of incredible on-device AI performance. I wanted to build something that proves this future:

🔥 An AI skill coach that runs locally on your phone and improves instantly based on your behavior — without ever sending your data to the cloud.

That idea became NanoSensei.

🤖 What it Does

NanoSensei is an offline-first AI coaching companion that analyzes micro-behaviors on your mobile device and gives real-time personalized feedback.

Key features

On-device inference (ExecuTorch / SQLite+vector embeddings stored locally) Voice and motion skill detection (device sensors) Instant feedback without cloud latency Zero data leaves the device — full privacy Graviton-powered backend for syncing anonymized session summaries Gamified progress scoring using AI embeddings

NanoSensei helps users learn any skill faster — from speaking confidently to improving motor-skill practice.

🛠️ How We Built It

Mobile (On-Device AI)

  • Built with React Native + ExecuTorch for mobile inference
  • Embedded a lightweight transformer-based skill evaluation model
  • Sensor fusion: accelerometer, microphone, and touch-event signals
  • Local embeddings stored in SQLite
  • No cloud inference required

Backend (ARM-Optimized)

Deployed on AWS Graviton (c7g.xlarge) for massive efficiency

  • FastAPI backend
  • Dockerized for ARM64
  • PostgreSQL with SQLModel for session storage
  • Simple sync API for user devices
  • Architecture tuned for low cost and high throughput

DevOps

  • Docker Compose (ARM64)
  • Secure SSH via PEM
  • CI/CD with GitHub Actions (optional future extension)

⚔️ Challenges We Ran Into

  • ExecuTorch model conversion required significant tuning
  • SQLModel raised errors with reserved names (metadata) — resolved by refactoring models
  • ARM64 Docker images needed custom builds
  • OneDrive Windows path permissions caused SSH key issues
  • Container kept restarting until deep inspection of logs
  • Rebuilding a backend fully compatible with ARM required several iterations

🏆 Accomplishments We're Proud Of

  • On-device LLM inference fully operational
  • Backend deployed 100% on ARM / Graviton, no x86 dependencies
  • Clean architecture: mobile → local AI → sync → Graviton backend
  • Achieved offline-first skill coaching, no cloud inference
  • Optimized Docker image size & startup time
  • Built a system that feels like the future of personal learning

📚 What We Learned

  • ARM-powered devices can run real AI workloads that used to require servers
  • ExecuTorch is a breakthrough for local inference
  • ARM Graviton simplifies cost and performance at scale
  • Careful naming in SQLModel prevents internal conflicts
  • Debugging containers via logs is essential with restarting loops
  • How to run end-to-end mobile + backend inference without ever relying on GPUs

🚀 What's Next for NanoSensei

  • Add more “skill packs” (speech, design gestures, mindfulness, drawing, fitness form)
  • Expand to real-time feedback with 30ms latency
  • Add WebGPU support for browser inference
  • Build a marketplace of custom skill-coaching AIs
  • Enable federated learning across devices
  • Offer SDK for other developers to embed NanoSensei in their apps
  • Launch on App Store + Play Store in 2026
Share this project:

Updates