Pocket Gull

Aerial Perspective for the Clinical Ocean
Real-time Care Plan Strategy & Live AI Consult Engine powered by Google Gemini

• ⚡ PocketGull WebGPU Zero-Egress Sovereign AI: 100% in-browser WebGPU hardware-accelerated clinical inference (HIPAA air-gapped).
• 🫀 PocketGull 3D WebGL Anatomy & Tri-Paradigm Triage: Three.js interactive 3D anatomy viewer with clickable organ nodes (Heart, Brain, Lungs, Liver, Spine).
• 💊 PocketGull ISMP Decimal Safety & CYP450 RxGuard: Instant prescription order safety auditor detecting 10-fold decimal errors and botanical interactions.
• 🕊️ PocketGull Clinical Intelligence Suite: Stepped-care triage acuity classifier and 3-Act Trajectory narrative generator.

• 🕊️ pocketgull-compass-2b: NIH/WHO Stepped-Care Triage, Socratic Health Literacy, and 3-Act Trajectories.
• 🕊️ pocketgull-sentinel-peft: Zero-Tolerance Emergency Red-Flag Interceptor & ISMP Decimal Safety Guard.
• 🕊️ pocketgull-scribe-soap: Zero-Egress Ambient Doctor-Patient SOAP & SBAR Encounter Encoder.
• 🕊️ pocketgull-tern-edge: Sub-45ms Ultra-Lightweight On-Device WebGPU / Mobile Edge Engine.
• 🕊️ pocketgull-albatross-multimodal: High-Capacity Tri-Paradigm Diagnostic & 3D WebGL Anatomy Integrator.
• 🕊️ pocketgull-rxguard-pgx: Pharmacogenomics & Botanical Supplement Interaction Screener.

# Register all models in 1 command
powershell -ExecutionPolicy Bypass -File dist/ollama/install_models.ps1   # Windows
bash dist/ollama/install_models.sh                                        # macOS / Linux

# Run anywhere
ollama run pocketgull-compass-2b

Pocket Gull gives practitioners the gull's eye view — the ability to rise above a turbulent sea of clinical data and see the clear, actionable patterns beneath.

It is a living clinical intelligence platform that synthesizes multimodal inputs — 3D spatial anatomy, bi-directional voice dictation, and standardized assessment instruments — into structured, evidence-grounded care strategies. Each care plan is examined simultaneously through Western Allopathic, Eastern TCM, and Ayurvedic paradigms, with AI reasoning powered by Google Gemini.

Pocket-Gull enforces human-first, clinical natural language chunking across text, speech, and live streaming:

• Prosodic & Respiratory Pacing: Audio streaming (AdkLiveService) buffers high-fidelity 24kHz/16kHz PCM audio in 32KB zero-copy frames (uint8ArrayToBase64), with 250–400ms natural conversational pause gating and client-side barge-in cancellation for authentic physician-patient turn-taking.
• Clinical Entity & Token Boundary Preservation: Slicers strictly prevent mid-token fragmentation of complex pharmacogenomic alleles (CYP2C19*17), blood pressure vitals (138/88 mmHg), lab values (eGFR 42 mL/min/1.73m²), and multi-word Latin botanical binomials (Withania somnifera).
• Defensive SSE Stream Reassembly: The streaming parser (GeminiProvider) dynamically reconstructs Server-Sent Events across packet boundaries, guaranteeing zero dropped tokens during network jitter while streaming clinical reports.
• Adaptive Cognitive Chunking: Automatically fragments dense medical consults into digestible, scan-friendly visual blocks (bulleted pearls, glassmorphic metric cards, and collapsible accordions) mapped to 5 health literacy personas.

• Procedural skeletal & organ viewer — Three.js geometry with severity-mapped particle systems
• Anatomical search with camera tracking — Fuzzy search bar that smoothly interpolates WebGL camera to targeted organs
• Raycast tooltips & data cards — Hover for paradigm badges and pain scores; click for slider input overlays
• Method of Loci memory palace — Anchors clinical consult nodes to 3D spatial coordinates for visual recall

Integrated interactive diagnostic tools accessible via the unified Clinical Tool Workbench:

Built-in validated clinical instruments integrated directly into patient state:

Specialized Google ADK agents maintaining real-time patient state context:

1. 🔭 Gulliver (Overview & Synthesis) — Holistic care plan strategy, multi-organ crosswalks, and timeline synthesis.
2. ⚡ Swoop (Interventions & Precision Dosing) — Targeted pharmacogenomic dosing, CPIC guidelines, and drug-botanical safety.
3. 🔦 Sentinel (Recovery Vigilance & Trends) — Continuous biomarker monitoring, Gompertz velocity tracking, and early warning signs.
4. 📖 Scribes (Patient Translation & Education) — Plain-language medical translation, health literacy bridging, and compassionate analogies.
5. ⚡ Skimmer (Flash AI Inference Backbone) — Sub-second edge triage, instant query routing, and real-time streaming.
6. 🚨 Samaritan (Emergency Override) — Offline BLS field guidance, 110 BPM CPR metronome, and first-responder EMT QR handoffs.

Dynamically reconfigures documentation, toolbars, and workflows for each user role:

• 🩺 Attending Physician / Clinician — High-efficiency CDS, 4-quadrant SOAP scribe, PGx RxGuard, and ICD-10 coding.
• 🏥 Resident / Fellow — Board exam differential radar, Socratic teaching pearls, and academic season tiering.
• 🔬 Clinical Researcher — $N$-of-1 crossover trial designer, NIH TrialFinder matcher, and FHIR ResearchStudy export.
• 🏛️ Hospital Executive / Health System Leader — QOF/HEDIS quality metrics, DiGA/FSE compliance, and privacy ROI.
• 🧑‍🤝‍🧑 Empowered Patient / Caregiver — Plain-language translation, SMS Compass health bridge, and life sovereignty goals.

• Classic Literary — Clean Caslon typography with optimal baseline grid leading.
• Bionic Speed Reading — Fixation point bolding for rapid optical scanning.
• Dyslexic Accessible — Specialized OpenDyslexic typeface with weighted bottom gravity.
• Audiobook Narrator — Web Speech API bi-directional voice narration.

Offline override mode for emergency field care:

• 110 BPM chest-compression metronome with BLS safety-gated AI
• FHIR-compliant EMT QR code serialization for first responder handoff
• Geo-Sentinel outbreak viewpoint deck — WHO, PAHO, and CDC surveillance modes
• Global telemetry suppression — all network calls disabled for offline triage

Adheres to Weniger, aber besser (less, but better) with WCAG 2.1 AA/AAA accessibility:

• 13+ curated themes — Rice Paper Washi, Raw Hemp, Carrara Marble, Dark Obsidian, Madame Curie Lab
• 4-level progressive disclosure — idle view → drill-down drawer → prescription state cycling → context menu
• Braun telemetry grid — monospace instrument panel headers with high-contrast metric readouts
• 44px+ touch targets — Fitts's Law compliant across all interactive elements

"Heavy DRM or server polling burns mobile battery and turns phones into pocket hand-warmers. Our lightweight mathematical verification consumes less energy than a single screen refresh, preserving all-day battery life for long hospital shifts."

• Sub-Microsecond Cryptographic Verification: Local SHA-256 salted hashing consumes $\approx 3\ \mu\text{J}$ (15,000x less power than waking a 5G/cellular modem for a remote API request), with zero flash memory wear ($0.000\text{ bytes written}$) and zero thermal degradation.
• Blinded Incognito Diagnostic Arena: Socratic active recall mystery cases for world leaders and scientific pioneers (Alexander, Caesar, Lincoln, Curie, Darwin, Ramanujan, Kahlo) with zero search-engine spoilers and 100% offline capability.
• Client-Side WASM & Web Workers: All Gompertz biomarker velocity models, Cohen's $d$ effect sizes, and Bayesian differentials execute purely on device, ensuring total patient privacy and uninterrupted reliability in hospital dead zones.

graph TB
    classDef cloud fill:#0f172a,stroke:#6366f1,stroke-width:2px,color:#f8fafc
    classDef ingestion fill:#0f172a,stroke:#38bdf8,stroke-width:2px,color:#f8fafc
    classDef hub fill:#18181b,stroke:#a855f7,stroke-width:3px,color:#fafafa
    classDef lenses fill:#0f172a,stroke:#10b981,stroke-width:2px,color:#f8fafc
    classDef foundation fill:#0f172a,stroke:#f59e0b,stroke-width:2px,color:#f8fafc

    subgraph Cloud ["Cloud & Backend"]
        CloudRun["Cloud Run"]
        Express["Express SSR + Proxy"]
        FastAPI["FastAPI Sidecar (ML)"]
        Vertex["Vertex AI / Gemini"]
    end

    subgraph Ingestion ["Ingestion & Portals"]
        Body3D["Three.js 3D Anatomy"]
        Voice["Voice Assistant"]
        Intake["Diagnostic Intake"]
    end

    subgraph Hub ["Central State & AI Orchestration"]
        State["PatientStateService\n(Signal Store)"]
        ADK["ADK InMemoryRunner\n(Multi-Agent)"]
        WebMCP["WebMCP Tool Catalog"]
    end

    subgraph Lenses ["Multi-Paradigm Lenses"]
        Western["Western Allopathic"]
        TCM["Eastern TCM"]
        Ayurvedic["Ayurvedic"]
        Ortho["Orthomolecular"]
    end

    subgraph Foundation ["Standards & Export"]
        FHIR["FHIR R4 Bundles"]
        Cache["Encrypted Offline Cache"]
        PubMed["PubMed Grounding"]
    end

    CloudRun --> Express
    Express <--> FastAPI
    Express <--> Vertex

    Body3D -->|Spatial Signals| State
    Voice -->|Audio Stream| ADK
    Intake -->|Vitals| State

    Express <-->|WS & REST| Hub

    State <--> ADK
    ADK <--> WebMCP

    Hub <--> Western & TCM & Ayurvedic & Ortho

    State --> FHIR & Cache
    ADK --> PubMed

    class CloudRun,Express,FastAPI,Vertex cloud
    class Body3D,Voice,Intake ingestion
    class State,ADK,WebMCP hub
    class Western,TCM,Ayurvedic,Ortho lenses
    class FHIR,Cache,PubMed foundation

• Node.js v24.x (strict — see .nvmrc)
• npm v10.x+
• Optional: Python 3.10+ for ML sidecar

# Clone
git clone https://github.com/pocketgull-app/pocketgull.git
cd pocketgull

# Install dependencies
npm install

# Start development server (Angular UI + Express SSR)
npm run dev

The app will be available at http://localhost:4200.

Create a .env.local file in the project root. See .env.example for the complete list.

pocketgull/
├── src/
│   ├── app.component.ts          # Root application component
│   ├── components/               # Standalone Angular components
│   ├── services/                 # Injectable services (AI, clinical, state)
│   │   ├── ai/                   # AI provider chain (Gemini, Hybrid, WebLLM)
│   │   ├── patient-state.service.ts
│   │   └── clinical-intelligence.service.ts
│   ├── server.ts                 # Express SSR server & API proxy
│   ├── server/                   # Server-side routes & services
│   ├── lib/                      # Firebase & DataConnect config
│   └── styles.css                # Global TailwindCSS stylesheet
├── companion-apps/
│   ├── avs-therapy/              # AVS Therapy companion (Node.js)
│   ├── patient_app/              # Flutter patient-facing app
│   └── provider_app/             # Flutter provider-facing app
├── pocketgull_api/               # Python FastAPI ML sidecar
├── docs/study/                   # Astro documentation site
├── e2e/                          # Playwright E2E tests
├── tests/                        # Vitest unit tests
├── scripts/                      # Build, deploy & security scripts
└── k8s/                          # Kubernetes manifests

Pocket Gull is built for clinical contexts with strict security posture:

• Zero PHI persistence — all patient state is transient (Angular Signals) or encrypted locally (Google Tink AEAD)
• Shift-Left Global Taint-Tracking — AST-level TypeScript Compiler API Data Flow Graph analyzer (scripts/taint-analysis-guard.mjs) verifying 0 un-sanitized source-to-sink flows
• Sentinel Security Guard — pre-commit egress domain whitelist enforcement and Shannon entropy secret scanning
• Polynomial ReDoS Prevention — static regular expression analyzer ensuring $O(N)$ non-backtracking parsing
• CodeQL 100% Remediation — hardened against SSRF, path traversal, prototype pollution, command injection, double-unescaping, and URL spoofing
• OpenSSF Scorecard: 10/10 — full supply chain security compliance
• SLSA Level 3 provenance — attested build artifacts
• CSP headers — strict Content Security Policy with nonce-based script isolation
• 1-click state purge — ephemeral data sovereignty per HIPAA Safe Harbor §164.514

See SECURITY.md for full vulnerability reporting policy.

Pocket Gull is architected as an empirical, peer-reviewed clinical intelligence testbed bridging engineering, computing, design, and health equity:

• Human-in-the-loop (HITL) — clinicians must validate AI output before archiving care plans
• Safety red-teaming — automated Vitest safety suite tests Gemini against adversarial prompts
• Evidence grounding — every recommendation anchored in PubMed literature with UKRIO citation formatting
• Skeptical epistemology — $p$-values against population baselines; Cochrane Risk of Bias assessments
• Evidence hierarchy tagging — recommendations tagged Level A (RCTs), Level B (Cohort), or Level C (Expert Consensus)

See RESPONSIBLE_AI.md for ethical principles.

All patient data serialized across API boundaries conforms to the FHIR R4 Bundle standard:

• 1-click FHIR R4 Bundle export (JSON)
• PDF care plan generation (jsPDF)
• Epic MyChart patient brief export portal
• SMART on FHIR OAuth 2.0 identity bridge
• CMS CPT 99453/99454/99457 RPM billing export

Deployed on Google Cloud Run targeting the gen-lang-client-0540208645 project:

npm run deploy

• Auto-scales to zero (minScale: 0) when idle
• Max 5 instances (maxInstances: 5)
• Artifact Registry 7-day cleanup policy
• GCS source bucket 7-day lifecycle policy

See CONTRIBUTING.md for guidelines. Key conventions:

• Conventional Commits: <type>(<scope>): <description> (72-char max subject)
• Standalone Components: No NgModules — Angular Signals over RxJS
• Pre-commit hooks: Husky enforces lint-staged, Sentinel guard, ReDoS audit, Global Taint-Tracking, and commit-msg format

Pocket Gull is an open-source medical intelligence ecosystem. You can back development directly on GitHub:

• 🌟 Community Backer ($10 – $25/mo): Supporter badge, listed in project release notes, and community vote on upcoming roadmap features.
• 🩺 Clinical Team & Lab ($100 – $500/mo): Priority issue triaging, early access to new AI models, and private roadmap advisory calls.
• 🏢 Enterprise & Health System ($1,000 – $5,000/mo): Dedicated SMART-on-FHIR connector support, custom on-premise deployments, and prominent corporate logo attribution.

If you reference Pocket Gull in research, please cite:

@software{gear_phil_2026_20647514,
  author    = {Gear, Phil},
  title     = {Pocket-Gull: Living Medical Intelligence Engine},
  month     = jul,
  year      = 2026,
  publisher = {Zenodo},
  version   = {v1.27.0},
  doi       = {10.5281/zenodo.20647514},
  url       = {https://doi.org/10.5281/zenodo.20647514}
}

Phil Gear — Lead Systems Architect & Creator
Member, IEEE (#101083329) · ACM Ambassador (#6179454) · Member, AIGA (#0513247) · Fellow, RSA (#8192283)
Arizona State University

_{© 2026 PocketGull LLC & Phillip Gear · MIT License}