docs: add ADR-007 for AI integration architecture and data-handling policy

docs/adrs/007-ai-integration.md

@@ -0,0 +1,49 @@
+# ADR-007: AI Integration Architecture and Data-Handling Policy
+
+**Date:** 2026-08-28
+**Status:** Proposed
+**Deciders:** Core team
+
+## Context
+
+AudioBlock is considering integrating AI-powered features such as recommendation engines, automated content tagging, and generative music tools. These capabilities require careful architectural planning to ensure:
+
+- User data (audio content, listening habits, wallet data) is handled securely
+- Artists retain control over how their content is used for AI training
+- AI services are loosely coupled to avoid vendor lock-in
+- The system remains compliant with privacy expectations
+
+## Decision
+
+Introduce AI capabilities through a dedicated **AI Service Layer** that acts as an adapter between the core backend and external or internal AI providers. The architecture follows these principles:
+
+1. **Opt-in only**: Artists must explicitly opt in before their content is used for AI training or inference beyond core platform features.
+2. **Data isolation**: AI training pipelines operate on a separate data store (or snapshot) that is populated only from opted-in content.
+3. **Provider abstraction**: All AI providers (recommendation, tagging, generation) are accessed through a common interface defined in `src/services/ai/`, enabling swap-in replacement.
+4. **No raw audio in prompts**: AI text-to-text features (e.g., lyric analysis) receive only the requested metadata, not raw audio streams.
+5. **Audit trail**: All AI data access is logged with artist ID, data type, and timestamp for compliance review.
+
+## Consequences
+
+### Positive
+- Artists retain full control; no silent data usage
+- Provider abstraction avoids lock-in to any single AI vendor
+- Audit trail supports future compliance needs (GDPR, etc.)
+- Separate data store limits blast radius of AI-related breaches
+
+### Negative / trade-offs
+- Additional infrastructure required for isolated AI data store
+- Opt-in flow adds friction for onboarding new artists to AI features
+- Provider abstraction layer adds modest code complexity
+
+### Neutral
+- Existing recommendation or analytics features are not affected until explicitly migrated
+- AI features are gated behind a feature flag (`AI_ENABLED`) for gradual rollout
+
+## Alternatives considered
+
+| Option | Why rejected |
+|--------|-------------|
+| Use artist content by default (opt-out) | Violates privacy expectations; legal risk |
+| Inline AI calls within existing services | Increases coupling; harder to audit data access |
+| Build all AI capabilities in-house | Beyond team scope; vendor solutions are mature for most use cases |