Inspiration

In sensitive industries like healthcare, law, and finance, professionals handle confidential documents containing Personally Identifiable Information (PII) and Protected Health Information (PHI). Standard cloud AI requires transmitting raw confidential text over public networks, risking HIPAA, GDPR, and corporate compliance violations. ScribeGuard AI was built to solve this dilemma through true hybrid intelligence: fast, sensitive drafting stays 100% on-device, and only pre-sanitized, high-order audits escalate to the cloud.

What it does

ScribeGuard AI is a privacy-first hybrid workspace that features:

  1. Client-Side Privacy Shield: Detects and masks sensitive emails, phone numbers, and national IDs in ~12ms directly in browser memory before any network request can occur.
  2. On-Device Gemini Nano Intelligence: Instant executive summarization and action-item extraction running locally on the user's device with zero latency and 100% offline capability.
  3. Firebase AI Logic Cloud Escalation: Seamlessly routes complex regulatory audits and large-context payloads to Cloud Gemini via the Firebase AI Logic SDK only after client-side sanitization.
  4. Live Telemetry & Latency Dashboard: Real-time KPI tracking of data kept on-device, PII entities shielded, and an interactive bar chart contrasting on-device speed (~18ms) against cloud roundtrips.

How we built it

ScribeGuard AI was built following Google's official Firebase AI Logic hybrid inference architecture:

  1. Client-side regex engine and token threshold evaluators running in the browser.
  2. Chrome Built-In AI / Gemini Nano (window.ai) integration for on-device prompt execution (≤1,800 tokens).
  3. Firebase AI Logic hybrid routing pattern with App Check verification and Remote Config threshold policies
  4. Responsive, high-trust light-mode user interface built with Tailwind CSS, Chart.js telemetry, and semantic HTML5, hosted on GitHub Pages.

Challenges we ran into

Structuring a clean architectural boundary that prevents sensitive data leakage while preserving the rich analytical quality of frontier cloud models, and ensuring that on-device token thresholds dynamically adjust based on document complexity.

Accomplishments that we're proud of

Delivering sub-20ms local inference with zero internet egress, establishing verifiable client-side PII sanitization in under 15 milliseconds, and crafting a clean, professional, distraction-free interface that regulated enterprise teams can immediately trust.

What's next for ScribeGuard AI

Adding local multi-modal document parsing (PDFs, clinical scans) using Chrome on-device vision models, and integrating Firebase Firestore for tamper-evident, client-encrypted compliance audit logs.

Built With

Share this project:

Updates

Submission history