Brand Guardian
Video ad compliance auditing

Every claim, checked against the rule that governs it.

Brand Guardian transcribes your video, reads its on-screen text, and holds every claim up against your own compliance rulebook — CAP Code, FTC guidance, whatever you've indexed. It never flags what it can't quote.

Audit a video → no signup — paste a link and go
vid_0982b848 — 0:32
Critical
00:29 …clinically proven to work in days, and
00:32 we guarantee results for every skin type.
Misleading Claims — absolute guarantees are prohibited without a substantiating clause. Matches CAP Code Rule 3.1, retrieved from your indexed rulebook.
CAP Code 3.1 exact quote required
2
LangGraph nodes — index, then audit
3
Rules retrieved per claim (top-k RAG)
0
Violations without a cited quote
What it does
Reads the whole ad
Full speech-to-text transcript and every frame of on-screen text (OCR), extracted via Azure Video Indexer — nothing is judged out of context.
Retrieves the governing rule
Your compliance PDFs are chunked and embedded once. Each claim is matched against the most relevant passages before judgment is made.
Won't flag what it can't quote
The auditor is instructed to cite an exact line from the transcript or OCR for every violation — and to leave puffery and taglines alone.
Two ways in
Paste a YouTube link for automatic download, or upload the file directly with title and description — the cloud-safe path.
How a video becomes a verdict
01
Ingest
The video is pulled in — downloaded from YouTube or accepted as a direct upload — and handed to Azure Video Indexer.
yt-dlp · Azure Video Indexer
02
Transcribe & read the screen
Azure extracts the full speech transcript and every on-screen text element, polling until processing completes.
speech-to-text · OCR
03
Retrieve the relevant rules
The transcript and OCR are embedded and matched against your indexed rulebook — only the top three most relevant passages are pulled in.
text-embedding-3-small · Azure AI Search
04
Judge, with citations required
GPT-4o weighs each claim against the retrieved rules under strict interpretation guidelines, and must quote its evidence to flag anything.
GPT-4o · temperature 0.0
05
Deliver the verdict
A structured PASS/FAIL report — every violation categorized, severity-tagged, and grounded in the exact line that triggered it.
FastAPI · async job polling
What gets checked
Critical Absolute or unsubstantiated claims — "guaranteed," "100% safe," "clinically proven" without support
Critical Missing mandatory disclosures — FTC, ASA, or platform-required disclaimers
Warning Composition vs. application ambiguity — is an effect attributed to the formula or the routine?
Warning Borderline claims that may need legal review but aren't a clear-cut breach
Ignored Puffery and brand taglines — exaggeration no reasonable viewer takes literally
Built with
LangGraph GPT-4o Azure Video Indexer Azure AI Search FastAPI OpenAI Embeddings LangSmith Docker