Content Intelligence
Governed AI capability families that turn raw media into structured, searchable, rights-aware context.
Governed AI capability families that turn raw media into structured, agent-addressable context. Each one ships as a governed tool reachable from any agent platform via CLI or MCP, metered through the gateway, traced end-to-end. Results land back in the Media Library with full provenance, so the next agent (and the next run) reasons against the enriched representation rather than re-deriving it.
The Content Intelligence layer is what makes media operationally addressable. It runs on top of the Media Library and writes structured results back into the bundle / version / provenance graph that the rest of the platform reasons against.
The capability families
Transcription & Captions
Turns audio into time-coded transcript and styled captions.
Whenever an agent needs spoken-word context — search, redaction, accessibility, derivative cuts.
Scene & Chapter Segmentation
Detects scene boundaries and chapter markers; surfaces structural metadata.
Editorial breakdowns, derivative-asset planning, smart navigation.
Brand Safety Classification
Classifies content against advertiser-relevant brand safety dimensions; produces structured advisories with confidence scores.
Advertising eligibility decisions, content-warning surfacing, vertical-specific safety policies.
Entity Extraction
Identifies named entities — people, places, organizations, products — across spoken and on-screen content; resolves them to canonical references.
Search, rights checking, cross-reference enrichment, knowledge-graph population.
Voice Synthesis (consent-bound)
Generates synthetic voiceover from text against rights-and-consent-bound persona libraries. Every generated clip is labeled in the Provenance Passport with the source persona and the consent record that authorized it.
ADR-style fills, multilingual narration, dynamic voice generation — always against consent-bound personas.
Creative Ingest
Walks an unknown input asset through the pipeline — probes, classifies, segments, and lands a structured representation back in the Media Library.
First-touch onboarding for new assets when the operator has no prior structure to bind to.
Each capability is callable independently. Each one composes with the others. A typical onboarding run might fire Creative Ingest, which in turn calls Transcription & Captions, Scene & Chapter Segmentation, Brand Safety Classification, and Entity Extraction in a single governed flow.
How the capabilities are governed
Every Content Intelligence call is a tool call through the gateway. That means:
Identity-bound. The platform knows which principal asked for the enrichment.
Org-scoped. Results land in the requesting organization's Media Library; cross-tenant access is not possible.
Metered. Each invocation is billed against the organization's credit ledger.
Audited. Every call produces an immutable audit row and a trace ID.
Schema-validated. Inputs are validated before the runtime sees them; malformed requests fail with a structured envelope.
There is no "internal call" path. The same enforcement chain applies whether a human invoked the capability through Pippa, an agent invoked it through MCP, or a MediaClaw composed it inside a multi-step run.
Where the results go
Content Intelligence does not produce ephemeral output. Each capability writes structured results back into the Media Library — at the appropriate bundle, version, or file boundary — so that later agents (and later runs) can reason against the enriched representation.
Specifically:
Transcription & Captions writes time-coded transcript artifacts and rendered caption tracks.
Scene & Chapter Segmentation writes segment markers that other tools and the MEDIA.md composer pick up.
Brand Safety Classification writes scored advisories at the bundle and segment level. All scores carry an explicit confidence; scores are advisory until calibrated against an organization's policy.
Entity Extraction writes resolved entity references back into the asset's metadata graph for search and cross-reference.
Voice Synthesis lands generated audio as a new file under the current version, with full provenance recorded.
Creative Ingest orchestrates the whole walkthrough and lands a populated bundle / version / files structure on first touch.
Every write is logged in the asset's Provenance Passport.
What you commit to when you turn this on
Per-call billing. Content Intelligence work is not free. Operators size their credit budget against expected enrichment volume.
Provenance records every step. AI-generated and AI-derived content is labeled in the Provenance Passport with the source capability and the model class used. This is not optional.
Brand-safety scores are advisory. They are signals for human policy, not autonomous gates. Operators decide how strict to be.
Where to go next
Looking for the Media Library these capabilities write into
Building a MediaClaw that composes intelligence calls
Studying decision memory and the learning loop
Connecting from your own harness
Last updated