+
+# Masker Voice Agents Hack
+
+Masker Voice Agents Hack explores a simple hypothesis: voice AI becomes more trustworthy when sensitive context can be detected, classified, and handled before it leaves the device.
+
+The goal of this project is to prototype an on-device privacy layer for voice interactions on mobile devices and wearables, using local intelligence to decide what should stay local, what should be masked, and what is safe to route onward.
+
+## Project Thesis
+
+If Masker can detect sensitive speech locally before it is forwarded upstream, voice AI becomes safer, faster, and easier to trust.
+
+That creates a compelling extension to the main product:
+
+- Mask PII before it reaches the LLM.
+- Decide locally whether a request should stay on-device, be masked, or be safely routed.
+- Build a voice-native privacy experience that feels real-time instead of bolted on.
+
+## Scope
+
+This repo is scoped as a prototype, not the full Masker platform.
+
+### In Scope
+
+- A voice-first demo built with Gemma 4 on Cactus.
+- On-device speech understanding for short live requests.
+- Detection of sensitive content in spoken input.
+- Local privacy policy decisions such as `local-only`, `masked-send`, and `safe-to-send`.
+- A clear demonstration of how on-device privacy logic complements the main Masker proxy product.
+
+### Out of Scope
+
+- Rebuilding the full `navi25/masker` platform in this repo.
+- Production-grade compliance guarantees.
+- Broad portal, dashboard, or customer management features.
+- Large integration surface area.
+- General-purpose assistant behavior.
+
+## Relationship To The Main Product
+
+The public `masker` repo answers:
+"How do we protect sensitive data before a voice agent sends it to an LLM?"
+
+This repo answers:
+"How early can we make that privacy decision, and can we make it directly on-device?"
+
+Together, the story becomes stronger:
+
+- `masker` is the production-facing masking and compliance layer.
+- `masker-voice-agents-hack` is the experimental edge intelligence layer.
+
+## Why Cactus + Gemma 4
+
+This prototype is built around Gemma 4 on Cactus because the stack supports the exact product question we want to test:
+
+- Can we do meaningful privacy-aware reasoning on-device?
+- Can a voice interaction still feel low-latency?
+- Can local classification improve trust before the request ever touches the network?
+
+## Prototype Flow
+
+1. A user speaks into the device.
+2. The request is transcribed and analyzed locally.
+3. Sensitive entities or risky context are detected.
+4. A local policy layer decides whether the request is `local-only`, `masked-send`, or `safe-to-send`.
+5. The result is either handled locally or forwarded as a masked / approved request into the broader Masker pipeline.
+
+## Why This Matters
+
+The current Masker product is already valuable because it protects voice-agent traffic before it reaches an LLM.
+
+This repo explores a sharper product wedge:
+privacy decisions that happen before sensitive speech even leaves the device.
+
+That matters for:
+
+- healthcare and compliance-sensitive workflows,
+- enterprise users speaking about internal projects or customer details,
+- and everyday users who want voice convenience without over-sharing by default.
+
+## Success Criteria
+
+This repo succeeds if it proves three things:
+
+1. On-device privacy detection is feasible in a voice workflow.
+2. The latency still feels conversational.
+3. The prototype expands the Masker story from "PII-safe voice agents" to "privacy-first voice AI from the edge inward."
+
+## Roadmap
+
+### Phase 1
+
+- Stand up the local voice loop
+- Detect basic PII classes
+- Implement route decisions
+- Show a working end-to-end demo
+
+### Phase 2
+
+- Improve sensitivity classification
+- Add user-visible privacy explanations
+- Measure latency and masking quality
+- Connect the prototype cleanly to the main Masker flow
+
+### Phase 3
+
+- Explore mobile and wearable packaging
+- Expand local policy customization
+- Use learnings to inform the core Masker platform roadmap
+
+## Repository Intent
+
+This repo should publicly read as:
+
+"The experimental on-device voice privacy layer for Masker."
+
+It exists to document the idea, prove the technical wedge, and show how edge reasoning can strengthen the broader Masker product.
+
+
+
+# π‘οΈ Masker
+### Programmable Compliance Layer for Cactus + Gemma Voice Agents
+
+> **Masker is a real-time privacy layer that makes local voice AI safe to ship.**
+It detects, transforms, and controls sensitive data before it reaches Gemmaβand filters what comes out.
+
+---
+
+## π¨ Why Masker (for THIS hackathon)
+
+Everyone here is building:
+- voice agents
+- using Cactus + Gemma
+
+But:
+
+- Users speak **PHI/PII** (SSN, medical info, phone numbers)
+- That data flows into:
+ - prompts
+ - logs
+ - responses
+
+> Even local AI is **not automatically private**
+
+- Logs leak
+- Responses leak
+- Agents mishandle sensitive data
+
+---
+
+## π‘ What Masker Does
+
+Masker sits **inside your Cactus + Gemma pipeline**:
+
+```
+User β Microphone β Cactus STT β Masker β Gemma β Masker β Cactus TTS β Speaker
+```
+
+---
+
+## π§ Core Capabilities
+
+- π Detect PHI/PII (SSN, phone, email, patient info)
+- π Transform (mask, tokenize, block)
+- π§ Apply policies (HIPAA-first)
+- π§Ύ Audit (trace what was filtered)
+- β‘ Runs locally (low latency, Rust-friendly)
+
+---
+
+## π§ Smart Privacy (No Compliance Toggles)
+
+> Donβt configure HIPAA vs GDPR. Masker figures it out.
+
+- Context-aware detection
+- Auto-applies safeguards
+- Works across healthcare + general PII
+
+---
+
+## π₯ HIPAA-First (Hackathon Scope)
+
+Masker ships with **3 simple policies**:
+
+### 1. Base Policy
+- Detect identifiers (SSN, phone, email)
+- Mask or tokenize
+- Allow symptoms
+
+---
+
+### 2. Logging Policy (Strict)
+- No PHI in logs
+- Store only masked or tokenized data
+
+---
+
+### 3. Clinical Policy (Context-aware)
+- Allow medical descriptions
+- Block direct identifiers
+
+---
+
+## βοΈ Integration (Cactus + Gemma)
+
+### π’ Option 1 β Auto Attach (recommended)
+
+```python
+from masker import auto_attach
+
+auto_attach()
+
+response = gemma.generate(user_input)
+```
+
+- No major code changes
+- Feels automatic
+
+---
+
+### π΅ Option 2 β Explicit Wrapper
+
+```python
+from masker import filter_input, filter_output
+
+safe_input = filter_input(user_input)
+response = gemma.generate(safe_input)
+safe_output = filter_output(response)
+```
+
+---
+
+### π‘ Option 3 β Local Interceptor (advanced)
+
+```
+App β Masker (local) β Gemma
+```
+
+---
+
+## π Gemma Backends
+
+`masker.gemma_wrapper` ships four backends behind one `GemmaBackend`
+protocol. `default_backend()` picks in this order:
+
+| Priority | Backend | Trigger | Use it for |
+|---------:|------------------------|--------------------------------------------|---------------------------------------|
+| 1 | `LocalCactusBackend` | `cactus` CLI on `PATH` | True on-device demo (laptop + GPU) |
+| 2 | `CactusCloudBackend` | `CACTUS_CLOUD_KEY` env var set | Cloud handoff w/o the cactus binary |
+| 3 | `GeminiCloudBackend` | `GEMINI_API_KEY` env var set | Direct Google API (no Cactus hop) |
+| 4 | `StubBackend` | nothing else available | CI / offline / unit tests |
+
+### Cactus Cloud quick start
+
+```bash
+cp platform/masker-library/.env.example .env
+# edit .env, paste your cactus_live_β¦ key into CACTUS_CLOUD_KEY
+set -a; source .env; set +a
+
+python -c "from masker.gemma_wrapper import default_backend; \
+ print(default_backend().generate('Reply with PONG.'))"
+# β PONG
+```
+
+`CactusCloudBackend` POSTs to the same `/api/v1/text` endpoint the
+`cactus` binary falls back to β uses stdlib `urllib`, no extra deps.
+SSL verification is off by default (matches the C++ FFI; the hosted
+endpoint uses a self-signed cert); set `CACTUS_CLOUD_STRICT_SSL=1` to
+flip it on, or override the host with `CACTUS_CLOUD_ENDPOINT`.
+
+---
+
+## π¬ Demo
+
+### Demo 1 β Cactus + Gemma Healthcare Agent
+
+- Run voice interaction
+- User speaks sensitive data
+- Show:
+ - detection
+ - transformation
+ - safe input to model
+ - safe output
+
+---
+
+### Demo 2 β Another Hackathon Agent
+
+- Plug Masker into another teamβs agent
+- Show same protection layer
+
+> **Goal:** prove Masker works across agents
+
+---
+
+## π Trust Models
+
+### Masker Managed (Default)
+- Fast setup
+- Masker handles encryption
+
+---
+
+### Customer Managed (AWS KMS)
+- Customer controls keys
+- Strong enterprise story
+
+---
+
+## 𧩠Architecture
+
+```
+Text Input
+ β
+Detection (PII / PHI)
+ β
+Policy Engine
+ β
+Transformation
+ β
+Gemma
+ β
+Output Filter
+```
+
+---
+
+# π 24-Hour Hackathon Plan
+
+## π― Final Goal
+
+- Masker integrated into:
+ - 1 primary Cactus + Gemma agent
+ - 1 external hackathon agent
+- Real-time masking working
+- HIPAA policies applied
+- Clear demo + story
+
+---
+
+## β±οΈ Execution Plan
+
+### π’ Stage 1 (Hours 0β4) β Core Engine
+
+**Goal:** working masking pipeline
+
+- [ ] Detect:
+ - SSN
+ - phone
+ - email
+- [ ] Transform:
+ - `[MASKED]`
+ - simple tokenization
+- [ ] Build:
+ - `filter_input()`
+ - `filter_output()`
+
+---
+
+### π΅ Stage 2 (Hours 4β8) β Gemma Wrapper
+
+**Goal:** attach to model
+
+- [ ] Build wrapper:
+ - `auto_attach()` OR `wrap(generate)`
+- [ ] Intercept input + output
+- [ ] Test with Gemma locally
+
+---
+
+### π‘ Stage 3 (Hours 8β14) β Policies + Trace UI
+
+- [ ] Define policies:
+ - Base
+ - Logging
+ - Clinical
+- [ ] Build minimal UI:
+ - input
+ - output
+ - trace
+
+---
+
+### π΄ Stage 4 (Hours 14β20) β Primary Agent Integration
+
+- [ ] Integrate with your main Cactus + Gemma agent
+- [ ] Run voice flow
+- [ ] Show real-time masking
+
+---
+
+### π£ Stage 5 (Hours 20β24) β External Agent Integration
+
+- [ ] Find 1 healthcare agent
+- [ ] Integrate Masker
+- [ ] Record fallback demo
+
+---
+
+## π§ Success Criteria
+
+- [ ] Real-time masking works
+- [ ] HIPAA policies applied
+- [ ] Demo polished
+- [ ] 1 external integration
+- [ ] Clear story
+
+---
+
+## π Final Pitch
+
+> βEveryone here is building voice agents with Cactus and Gemma.
+Masker is the layer you add before you shipβto make them safe.β
+
+---
+
+## π₯ One-liner
+
+> **Masker is a real-time privacy layer for local voice agents.**
diff --git a/README.md b/README.md
index 54e31c3e..dc3d2a3b 100644
--- a/README.md
+++ b/README.md
@@ -41,4 +41,16 @@ Prizes per special track:
## Next steps
1. Read Cactus docs carefully: [Link](https://docs.cactuscompute.com/latest/)
2. Read Gemma 4 on Cactus walkthrough carefully: [Link](https://docs.cactuscompute.com/latest/blog/gemma4/)
-3. Cactus & DeepMind team would be available on-site.
\ No newline at end of file
+3. Cactus & DeepMind team would be available on-site.
+
+## Repository Layout
+
+```text
+platform/
+ masker-core/ # Rust runtime and CLI
+ masker-admin/ # demo and trace UI
+ masker-library/ # Python SDK, integration package, tests, and env samples
+```
+
+This keeps the product surfaces grouped together instead of splitting the repo
+by language-specific top-level folders.
diff --git a/aurora/.gitignore b/aurora/.gitignore
new file mode 100644
index 00000000..5983e210
--- /dev/null
+++ b/aurora/.gitignore
@@ -0,0 +1,12 @@
+node_modules/
+.expo/
+dist/
+ios/Pods/
+ios/build/
+android/build/
+android/app/build/
+android/.gradle/
+*.log
+.DS_Store
+.env
+.env.local
diff --git a/aurora/App.tsx b/aurora/App.tsx
new file mode 100644
index 00000000..d29cef9d
--- /dev/null
+++ b/aurora/App.tsx
@@ -0,0 +1,37 @@
+import 'react-native-reanimated';
+import React, { useState } from 'react';
+import { StatusBar } from 'expo-status-bar';
+import { GestureHandlerRootView } from 'react-native-gesture-handler';
+import { SafeAreaProvider } from 'react-native-safe-area-context';
+
+import { ConversationScreen } from '@/screens/ConversationScreen';
+import { OnboardingScreen } from '@/screens/OnboardingScreen';
+import { DisclaimerModal } from '@/components/DisclaimerModal';
+
+type Stage = 'onboarding' | 'disclaimer' | 'conversation';
+
+export default function App() {
+ const [stage, setStage] = useStateIssue credentials for SDKs and services. Scope by environment and endpoint.
+| Label | +Key Prefix | +Permissions | +Environment | +Last Used | +Status | +Actions | +
|---|---|---|---|---|---|---|
| {k.label} | +
+
+ {k.prefix}β’β’β’β’
+
+ |
+
+
+ {k.permissions.map((p) => (
+
+ {PERMISSION_LABELS[p] ?? p}
+
+ ))}
+
+ |
+ + + {k.environment} + + | +{formatLastUsed(k.lastUsed)} | +
+
+
+
+
+ |
+
Generate downloadable evidence by use case, framework, and date range.
+| Report | +Use Case | +Date Range | +Sessions | +Generated By | +Status | +Download | +
|---|---|---|---|---|---|---|
|
+
+
+ |
+ {r.useCase} | +{r.dateRange} | +{r.sessionCount > 0 ? r.sessionCount.toLocaleString() : "β"} | +{r.generatedBy} | +
+ {r.status === "ready" ? (
+
+ {r.formats.map((f) => (
+
+ ))}
+
+ ) : (
+ β
+ )}
+ |
+
+ Tell the copilot what you are building. It will propose a structured compliance configuration you can review, test, and publish. +
++ {error.message || "An unexpected error occurred. Try refreshing the page."} +
+ {error.digest && ( +Error ID: {error.digest}
+ )} +Keys and secrets, without operational chaos.
+| Key Alias | +Scope | +Region | +Created | +Rotation | +Last Rotated | +Status | +Actions | +
|---|---|---|---|---|---|---|---|
| {k.alias} | +{k.scope} | +{k.region} | +{k.created} | +{k.rotationCadence} | +{k.lastRotated} | +
+
+
+
+
+
+ |
+
+ {new Date().toLocaleDateString("en-US", { month: "short", day: "numeric", year: "numeric" })} Β· Production Β· Acme Health +
+{inc.summary}
+ + {inc.id} + +| Report | +Use Case | +Status | +Download | +
|---|---|---|---|
| {r.name} | +{r.useCase} | ++ {r.status === "ready" && ( + + + + )} + | +
Manage rules, versions, and entity-level actions.
+| Name | +Framework | +Scope | +Version | +Status | +Last Updated | ++ |
|---|---|---|---|---|---|---|
| {p.name} | ++ {p.framework} + | +{p.scope} | +v{p.version} | +{p.updatedAt} | ++ + | +
| Session ID | +Time | +Channel | +Use Case | +Policy | +Status | +Entities | +Risk | +Duration | ++ |
|---|---|---|---|---|---|---|---|---|---|
|
+
+
+
+ No sessions match your filters + + |
+ |||||||||
| {s.id} | +{formatTs(s.timestamp)} | +
+ {s.channel === "voice"
+ ? |
+ {s.useCase} | +{s.policyVersion} | +{s.entitiesDetected} | +{s.duration} | +
+ |
+
+ Click any row to inspect the transcript diff, entity spans, and audit trail. +
+Workspace configuration, team access, integrations, and retention.
++ {n.text} +
+