PII-Zero: Zero-Trust Clinical & Legal Hybrid AI Airlock

Clinicians dictating hospital discharge summaries and attorneys reviewing privileged M&A schedules need the deep reasoning power of fro

Comment

PII-Zero: Zero-Trust Clinical & Legal Hybrid AI Airlock

Inspiration

Clinicians dictating hospital discharge summaries and attorneys reviewing privileged M&A schedules need the deep reasoning power of frontier cloud models (Gemini 2.5 Flash), but strict regulatory mandates (HIPAA Safe Harbor 45 CFR 164.514(b)(2) and FRE 502 Attorney-Client Privilege) forbid transmitting raw patient names, SSNs, Medical Record Numbers (MRNs), docket numbers, or confidential settlement reserves to external cloud endpoints.

Neither cloud-only AI nor on-device-only AI can solve this alone:

  • Cloud-only AI leaks raw PII/PHI across the network boundary before redaction can happen.
  • On-device-only AI keeps data local, but compact edge models lack the deep clinical differential and multi-clause legal reasoning capacity of Gemini 2.5 Flash.

PII-Zero resolves this paradox by turning on-device Gemini Nano 4 into a hardware-local Zero-Trust Privacy Airlock and Cloud Gemini 2.5 Flash into a Zero-Knowledge Deep Reasoning Engine, orchestrated via the Firebase AI Logic Hybrid SDK.

What It Does

PII-Zero executes a 4-stage cryptographic and hybrid inference pipeline across Web (https://piizero.gilbertcv.com), Chrome Extension (Manifest V3 Side Panel), and Android (Kotlin + Jetpack Compose with Galaxy Z Fold7 FoldingFeature support):

  1. Stage 1 — On-Device HIPAA 18 & Legal Privilege Airlock (InferenceMode.PREFER_ON_DEVICE / ONLY_ON_DEVICE): Scans raw clinical notes or legal contracts locally on the client device, replacing all 18 HIPAA Safe Harbor identifier classes and privileged legal entities with deterministic typed tokens ([PII_PERSON_1], [PII_SSN_1], [PII_MRN_1], [PII_DOCKET_1], [PII_SETTLEMENT_1]). The mapping between tokens and raw values is stored exclusively in a Client-Only Reversible Token Vault in volatile local memory.
  2. Stage 2 — Cryptographic Zero-PII Gatekeeper (SHA-256 Attestation): Before any network request can be dispatched, an independent verifier inspects the sanitized payload to prove residualCount === 0 and generates a tamper-evident SHA-256 attestation digest of the exact outbound bytes. If even a single unredacted identifier remains, outbound cloud dispatch is hard-blocked.
  3. Stage 3 — Cloud Gemini 2.5 Flash Deep Synthesis (firebase/ai & com.google.firebase:firebase-ai): The sanitized text (containing only [PII_] tokens) is analyzed by **Gemini 2.5 Flash*, which produces a structured clinical handoff brief or legal risk audit while preserving the placeholder tokens verbatim.
  4. Stage 4 — Client-Side Token Re-Hydration: When the cloud response returns to the device, PII-Zero re-hydrates [PII_PERSON_1] -> Elena Rostova and [PII_MRN_1] -> MRN-994821A on-screen for the authorized clinician or attorney—meaning the cloud model reasoned over the full case context without ever seeing a single byte of raw PII.

How We Built It

  • Firebase AI Logic Hybrid Inference SDK:
    • Web & Chrome Extension: Built on firebase/ai (getAI, getGenerativeModel, GoogleAIBackend, InferenceMode.PREFER_ON_DEVICE, InferenceMode.ONLY_ON_DEVICE, InferenceMode.ONLY_IN_CLOUD) paired with Chrome Built-in AI (self.ai.languageModel / LanguageModel Gemini Nano 4 bridge).
    • Android App: Built in Kotlin with com.google.firebase:firebase-ai (Firebase.ai(backend = GenerativeBackend.googleAI()).generativeModel(...)) and FirebaseAppCheck (DebugAppCheckProviderFactory), plus a deterministic local HIPAA/Legal airlock fallback so the APK runs out-of-the-box offline.
  • Chrome Extension (Manifest V3):
    • Uses sidePanel, contextMenus, activeTab, storage, and scripting permissions.
    • Right-clicking any selection on an EHR or contract page triggers "Scrub PII & Analyze via PII-Zero Hybrid Airlock", and the content script supports 1-Click Inline Replacement of page form fields with redacted [PII_*] tokens.
  • Foldable-First Android UI (Jetpack Compose + androidx.window:window):
    • Uses WindowInfoTracker and FoldingFeature to detect device posture in real time. On an unfolded Galaxy Z Fold7 or tablet, it renders a 2-Pane Hinge-Aware Split Ledger (24dp hinge clearance). On a compact cover screen, it adapts to a 3-Tab Single-Pane Ledger.
  • Anti-Slop Clinical Ledger Design System (designer-mcp):
    • Grounded in unslop-original-web-baseline, unslop-audit-checklist, recipe-analytics-dashboard, and github-mobile-compose-reply: #111C26 slate canvas, #182734 ruled panels, 1px solid #263849 hairline borders, #8DD6C7 safe attestation teal, #F2CB85 amber token badges, and tabular-nums monospace alignment.

Challenges We Ran Into

  • Preserving Coreference Across Redaction: Generic [REDACTED] masks destroy coreference resolution across multi-party contracts or clinical consult notes. By assigning deterministic entity-indexed tokens ([PII_PERSON_1], [PII_PERSON_2], [PII_MRN_1]), Gemini 2.5 Flash preserves entity relationships accurately and allows lossless local re-hydration.
  • Cryptographic Gatekeeping Before Socket Dispatch: We added a dedicated post-scrub verifier that re-evaluates all 18 HIPAA identifier patterns and hashes the sanitized payload with SHA-256 before permitting any cloud call.
  • Adaptive Foldable Continuity: Coordinating layout transitions between a 360px Chrome Side Panel, a narrow smartphone cover screen, and an unfolded Galaxy Z Fold7 2-pane split without visual clutter.

Accomplishments That We Are Proud Of

  • Full 3-Platform Delivery Targeting the Grand Prize: Web App + Chrome MV3 Side-Panel Extension + Kotlin Compose Foldable Android App.
  • End-to-End Automated Verification: Unit and build verification passing across Web (vitest + vite build), Chrome Extension (node --test + esbuild + zip), and Android (testDebugUnitTest + assembleDebug).

Try It Links

Built With

Share this project:

Updates