AI Integration

:::info Source Sourced from services/media-service/AI_INTEGRATION.md in the documentation repo. :::

1. AI Capabilities

Capability	Prompt	Classification
Image generation	media.image.generate	Limited-risk
TTS (text-to-speech)	media.audio.tts	Limited-risk
Auto-captioning (STT)	media.stt.caption	Limited-risk
Transcript generation	media.stt.transcript	Limited-risk
Image alt-text	media.image.alt_text	Limited-risk
Content safety pre-scan	media.safety.classify	Safety
Image-to-image edit (M5+)	media.image.edit	Limited-risk

All via AIClient port with provenance.

2. Safety Pipeline

• Pre-generate: prompt moderated by ai-gateway.
• Post-generate: image moderated (NSFW, violence, CSAM).
• Hit → quarantine + law enforcement reporting if CSAM.

3. Provenance

Every AI artifact has aiProvenance:

• Model + prompt ID + version.
• Traces back to requesting user + decision ID.
• Visible in UI as "AI-generated" badge.

4. Cost Controls

• Per-tenant AI budget enforced at ai-gateway.
• Image gen ~$0.02 per image.
• STT ~$0.006 per minute.
• Budget UI alerts at 80%.

5. Local vs Cloud

• Image generation: cloud (quality).
• TTS: cloud preferred; local for quick preview (lower quality).
• STT: cloud + optional on-device for privacy-sensitive tenants.

6. Data Privacy

• Uploaded images → AI for alt-text: with tenant consent; redaction applied.
• noTrain flag verified on all providers.
• HIPAA tenants: restricted providers with BAA.

7. Caching

• Same (prompt hash, model, size) → cache 30 days.
• STT of same (asset sha256, model, language) → cache indefinitely (deterministic).

8. Right to Explanation

• AI-generated captions show confidence score.
• Alt-text shows what the model "saw" + suggests review.

9. Bias & Accessibility

• Caption quality eval quarterly (accuracy on diverse speakers).
• Image gen bias eval on demographic prompts.