The Media Customs

Start localizing your video content for every geography and culture

An agentic ad-clearance crew. It watches your commercial once, then judges it in parallel against every market you ship to, on both counts: the law, and the cultural ground it has to land on. Gesture, modesty, religious symbols, superstition, humour, politics. It builds its own Grafana instrument panel as it goes, and re-renders the shots that fail.

Come in as

I'm a Devpost judge Straight to the runs already cleared, with the evidence behind every finding. No password: reading is open Show me around The whole product in a few screens, then a guided walk through the live console. About two minutes I'm a curious visitor Bring a commercial and watch it clear customs, market by market. No password

Watch a fix land three markets · three statutes · three fixes the verifier signed off

FR · Loi Évin · alcohol

before

after

A toast in red wine, which French law keeps off the air. Re-rendered: the wine is juice, the bottle a carafe.

EU · AVMSD art. 9 · tobacco

before

after

A cigarette at a café table, banned outright by the EU directive. Re-rendered: an espresso, and no smoke.

AE · Decree-Law 55/2023 · modesty

before

after

A hemline that keeps the spot out of Gulf rotation. Re-rendered at the knee, same fabric, same walk.

The whole thing, in a few minutes watch it on YouTube

Why this exists

Global brands ship one commercial into dozens of markets and check maybe six of them. France bans alcohol advertising under Loi Évin. Quebec bans advertising aimed at children under 13. Thailand's lèse-majesté law makes a careless royal reference a criminal matter. The UK requires every ad to pass Clearcast before it airs.

And the unregulated half costs more than the regulated half. Pepsi and Kendall Jenner. Dolce & Gabbana in China. H&M's hoodie. None of those broke a statute. Every one of them was caught by the public instead of by a process. This is the process.

What happens when you upload a video observe once, judge many

The architectural decision everything else follows from. One multimodal analyst watches the film a single time and writes down what it sees: neutral, timecoded, no verdicts. Then one adjudicator per jurisdiction judges that same fact sheet against its own rulebook, in parallel.

1 You hand it a master: an MP4 or a MOV, up to two minutes.
2 Ingest cuts it into shots and pulls the audio and a transcript.
3 The analyst watches it once and writes timecoded, neutral observations. They go to Loki as they are made.
4 That one fact sheet goes to every market at once. An adjudicator per jurisdiction judges it against its own rulebook and grounds its citation against a live source.
5 Each finding pushes a severity metric to Mimir on the film's own clock, and an annotation onto the run's dashboard.
6 The publisher agent builds the dashboards and the alert rules itself, through the Grafana MCP server.
7 A blocking finding trips its alert rule, and the webhook wakes the remediator. This is the arrow that makes Grafana a participant: the fix starts because a chart said so.
8 The verifier re-observes the new footage with the same instrument. The metric drops, Grafana resolves its own alert, and the localized master and the certificate are yours.

01 Ingest Shots, audio and a transcript. The film is broken into the pieces a regulator would argue about.

02 Analyst One pass, 18 dimensions, and verdicts forbidden. “A woman raises a glass of red wine” is an observation. “This breaks Loi Évin” is not.

03 Adjudicators One per jurisdiction, all at once, 98 jurisdictions in this build, each grounding its citation against a live source. A finding is always observation × rule × citation.

04 Guard A pure rule layer with no model in it. Where a market's rule targets who appears in your ad, it refuses to auto-fix and hands you the statute.

05 Publisher Metrics, logs, annotations and dashboards, written into Grafana by the crew itself over MCP.

06 Remediator Woken by a Grafana alert. Re-letters, re-props, re-voices or regenerates the failing span, priced in euro before you press.

Two ways in: click it, or ask it agent mode

Everything above is a screen you can click. It is also a tool the agent in front of it can call, and those are the same functions, not a second implementation: an agent that answers “what is blocked in France” out of its own summary of a summary is a demo, and this one answers it by reading the store the market room reads. Ask for something nobody wrote a screen for and it writes the LogQL, builds the dashboard and hands you the URL.

1 It drives the product Start a clearance, add markets, price a fix, open a frame, search the corpus, compose a chart. 11 tools, every one of them a thing this console already does.

2 It answers with the evidence open The reply is on the left and what it looked at is on the right: the actual board, room, frame or panel, not a description of one.

3 It obeys the same refusals The Guard still blocks a protected-basis fix, scope still refuses a patch it cannot support, and the day's budget still governs Veo. No sentence talks its way past a rule.

Open agent mode

Grafana is a participant, not a picture

1 The crew writes into it Seven dashboards, four metrics, three kinds of log line and an annotation on every finding. The publisher is an agent with real MCP tools, and the dashboard description is prose it wrote itself.

2 It triggers the crew An alert rule fires a webhook that wakes the remediator. Grafana sits upstream of the work rather than in a report after it.

3 The crew reads it back Every chart in this console is a Loki or Mimir query. Ask the agent a question nobody wrote a screen for and it writes the LogQL itself.

Built with

Agent BuilderAgent Development Kit

Built for the Agentic Cinema hackathon, Grafana Labs track. Gemini 3.7 and Veo 3.1 on Vertex AI, Grafana Cloud. Source on GitHub.