Zeda

AI-powered dual defense against fake content and hidden trackers. Verify screenshots with a Veracity Score and block identity profiling in real time-all in one digital security hub.

Comment

Inspiration

We built Zeda because misinformation and phishing are now mixed into everyday workflows: emails, social posts, and shared screenshots. During the hackathon, we wanted a tool that helps people pause before trusting content, especially when messages use urgency, fear, or fake authority.

What it does

Zeda has two modes:

  • Verify: checks whether a claim is supported by evidence.
  • Protect: detects scam/privacy risk patterns in text, URLs, and images.

Users can submit raw text, a URL, or a screenshot. Zeda extracts content, analyzes it, assigns a score, explains why, and returns a report with suspicious parts and source/evidence details when available.

In the extension, users can run the same flow directly on any page from a slide-in sidebar, including screenshot capture (full view or selected area) and paste-from-clipboard.

How we built it

We built a full-stack pipeline:

  • Frontend web app: React + Vite UI with mode-based flow (Verify/Protect), upload/input handling, and results rendering.
  • Backend API: Node.js + Express endpoints for /api/ingest, /api/verify, and /api/protect.
  • Content ingestion:
    • OCR for image inputs.
    • URL scraping/normalization for web inputs.
    • Text normalization for raw inputs.
  • Analysis layer:
    • Claim/risk analysis with LLM-backed prompts.
    • Heuristic scoring + structured output normalization.
  • Browser extension (Manifest V3):
    • Background service worker for secure API calls.
    • Content script + Shadow DOM sidebar UI.
    • Page capture workflows and quick analysis actions.
  • Performance UX:
    • Per-step timing in UI: [ T_{\text{total}} = T_{\text{extract}} + T_{\text{analyze}} + T_{\text{report}} ]
    • Users can see where latency is spent.

Challenges we ran into

  • Large screenshot payloads causing 413 Payload Too Large.
  • OCR reliability on complex/dark screenshots and noisy images.
  • Timeout behavior and slow end-to-end analysis under hackathon constraints.
  • Score calibration drift (legit content too risky, scams too low).
  • Frontend result-shape mismatches causing rendering errors.
  • MV3 extension permissions and capture constraints (activeTab, host permissions).
  • Syncing branch changes quickly while avoiding merge conflicts.

Accomplishments that we're proud of

  • Delivered a working multimodal trust-checking product in hackathon time.
  • Shipped both a web app and a browser extension with consistent behavior.
  • Built a practical Verify/Protect workflow with interpretable scoring and reasons.
  • Added screenshot and clipboard-driven analysis for real browsing contexts.
  • Improved transparency with step-level timing in analysis progress.
  • Stabilized UX and scoring after iterative debugging and calibration.

What we learned

  • Reliability comes from fallbacks + normalization, not just model quality.
  • Scoring systems must be calibrated with adversarial and benign examples.
  • MV3 extension architecture requires careful separation of UI and privileged logic.
  • Real user trust improves when the system explains both score and evidence path.
  • Tight team coordination and disciplined branch hygiene are critical under time pressure.

What's next for Zeda

  • Stronger grounding and citation quality for Verify results.
  • Better phishing/domain intelligence and richer Protect signals.
  • Faster inference path with caching and smarter parallelization.
  • Improved OCR pipeline for difficult screenshots.
  • Persistent report history and collaboration/sharing features.
  • Publish-ready extension hardening and store deployment.
  • Continuous evaluation suite to track false positives/false negatives over time.

Built With

Share this project:

Updates