Aditya Pai.
Résumé
← All work

02 / On-Device AI / Privacy Engineering

Little Helper

Privacy-first, on-device child safety

Role
Software engineering
Period
2026
Built with
Swift · Python · Core ML · Electron · Supabase · WebRTC
Intelligence stays hereCONCEPT STUDY
VisionSpeechTextOCR
On device first. Private by design.
Conceptual engineering diagram · Illustrative, not a product screenshot or measured plot.
per inference check
3.1 ms
peer-to-peer monitoring
~4 FPS

The problem

A child-safety system needs to understand activity across multiple input types without making continuous cloud uploads the default. It also has to respond quickly enough to work alongside everyday applications.

I built an offline macOS safety engine spanning vision, OCR, typed text, and speech, with a separate family-management and live-monitoring path.

Local intelligence first

ScreenCaptureKit supplies screen capture on macOS. Core ML supports local inference; the surrounding Swift and Python components coordinate detection across input types. Electron provides the desktop application layer.

Trust and access

ILLUSTRATIVE ARCHITECTURELocal inference. Separate remote paths.
  1. 01
    Device inputsScreenCaptureKit, OCR, typed text, and speech
  2. 02
    Local classificationOffline Core ML safety engine
Optional escalationPrivacy-preserving classification upload to Claude
Live monitoringEncrypted peer-to-peer WebRTC, with TURN relay when needed

Family-scoped PostgreSQL RLS and hashed device tokens control access. The live video path is separate from classification escalation.

The backend uses hashed device tokens and family-scoped PostgreSQL row-level security. Classification escalation to Claude uses privacy-preserving uploads; access control is enforced at the database boundary.

A separate live-monitoring path

I rebuilt screen monitoring around peer-to-peer WebRTC, Electron, and Cloudflare TURN. The implementation improved observed throughput from approximately 0.5 to 4 FPS. TURN can relay encrypted media when direct connectivity fails; the application backend does not store the video stream.

Results

  • Inference latency reduced from 14 ms to 3.1 ms per check.
  • Vision, OCR, text, and speech supported in an offline macOS engine.
  • Live-screen throughput improved to approximately 4 FPS in the project setup.
  • Family-scoped access and privacy-preserving escalation implemented.

Engineering takeaway

Fast local decisions and remote visibility have different constraints. Keeping those paths separate makes it possible to optimize responsiveness without turning every inference into a network request.