Javier Stewart
Back to projects

Active

Lens

Privacy-first phishing copilot that learns locally and explains decisions.

Lens is a Chrome extension (Manifest V3) built for on-device phishing detection across email and document workflows. It runs content scanning locally, combines language/link/context signals, and explains verdicts with confidence and reason cues so decisions are transparent.

It includes train/protect modes, trust controls, and advanced diagnostics while keeping private content on-device. The architecture uses content scripts + a service worker, local storage-based learning, and security guardrails to keep feedback useful without poisoning the model.

Highlights

  • Explainable risk panel with confidence bands and evidence cues
  • Advanced diagnostics for signal breakdown, combos, and scan latency
  • Trust controls: Yes/No feedback, trust sender/link, and reversible actions
  • Safety interstitial before opening suspicious links
  • Local-first learning model with anti-poisoning guardrails
  • Protect vs Train modes with sensitivity and policy controls

Screenshots

Lens risk panel in Gmail showing medium risk with reasons and trust controls.
Risk summary panel with explainable cues and user trust controls.
Lens advanced diagnostics showing score components, confidence, and triggered rules.
Advanced diagnostics for score transparency, confidence, and triggered rules.
Lens popup controls with protect and train modes plus sensitivity settings.
Popup controls for mode, sensitivity, and local privacy-focused configuration.

Tech Stack

  • Chrome Extension (MV3)
  • Service Worker
  • Content Scripts
  • Chrome Storage API
  • Local Learning Guardrails
  • Explainable Risk Scoring
  • Playwright E2E