Skip to content
RTK23-devPublic

About

Advertising for one brand: evidence, the ads, and what happened next. Paid ads are created paused.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

1,593 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Meridian

Meridian turns a brand's research into original content batches. It gathers evidence, helps choose what to make, plans and produces video and image variants, checks them, and leaves a person to review before anything is published. Publishing, paid ads and analytics are follow-on workflows, not prerequisites to making content.

It never invents a result. A step with no key, no account, or no bill says so and stores nothing. A person reviews what is made before it goes out, and paid ads are always created paused.

What it is for

The intended primary job is to use research and reference examples to choose a repeatable video format, generate many original variants, review them, and export the finished work. Meridian also has brand, publishing, ads and measurement tools, but the current interface does not yet guide this end-to-end job in one place.

How it works

There are two ways to start, and they share one backend: the same brands, evidence, plans, assets, runs, reviews, publishing receipts, measurements and learning.

  • Research-led. Batch gathers and grades evidence about the niche, proposes an angle, plans, produces, checks, and asks for approval.
  • Direct. You bring a brief, a reference video, or your own upload, and the Content page plans, generates, revises and reviews it.

These are still separate entry paths in the interface. Research collected in Market is not yet presented together with a reference-led video batch in one guided project.

The loop, in either case:

  1. Brand. Positioning, customers, tone, claims, and products.
  2. Evidence. Pages, ads, and uploads the brand is allowed to use.
  3. Decision. The decision engine answers a fixed set of questions from that evidence. It does not invent the evidence.
  4. Brief. An approved decision becomes a creative brief.
  5. Make. Images go to the image provider configured for the workspace. Video goes through OpenRouter with the model chosen in Settings, Models, or a manual cloud folder. The direct Gemini Omni and Higgsfield routes are switched off in this version. Hypit is not connected to production routing: it refuses every job until its runtime is connected and licensed (docs/HYPIT_LICENSE.md). If the chosen provider is not connected, the job says so and nothing is sent.
  6. Check. Claims, brand rules, rights, and a human review. A missing piece of evidence goes to review. It is never passed.
  7. Publish. Posting to a connected Facebook Page works: each creative is claimed per Page before the post is sent, so it is not posted twice, and an unknown outcome is held as uncertain until a person checks Facebook and confirms. Instagram, YouTube and TikTok posting is switched off in this version: export the video and post it yourself. Paid ads are created paused, and Meridian has no button that turns them on. A delivered video can also be repurposed into reviewed variations, and each step (library, upload, paused ad, Facebook post, metrics) is a separate action; see docs/FACTORY.md.
  8. Learn. Measured results are stored only when they were observed, and the next brief can change.

The screens

Screen What it is for
Setup guide A short path: create a brand, choose a decision engine and save its key. OpenRouter shows key saved, test passed and engine active as separate steps. Optional connections sit behind "Open this step"
Market Paste one public page URL, with no setup, and see what was read, what is new, what was blocked and what to do next. Searching by phrase is a separate, optional step with your own developer key
Content One page per brand. Create and Batch are the first sections; Ideas, Review, Library, Publish and Results follow. Batch opens on Video batches: use an uploaded reference video and attached Market evidence in a reusable draft, create 2–20 original draft variations with a bounded total budget, then explicitly start the batch and review/export each video. A batch takes its style from a cookbook (a saved recipe for one kind of video, see docs/COOKBOOKS.md); the style control filters the cookbooks. The research-led run builder remains an advanced option. Live provider generation and owner acceptance are not yet verified.
Billing Every money limit in one place: the workspace monthly budget, each brand's budget, the factory's daily and total caps, the measure ceiling and the OpenRouter decision cap. Nothing is set by default, and paid work is refused until its limit is set. Only an admin can change a limit, and each change is audited
Integrations Every connection, with what it needs and its state
Jobs & health Queued work, the worker and the scheduler
Settings The workspace, its members, scoring weights, alerts, learning, and the uncertain model-call holds

What it will not do

These rules hold across the code, and tests enforce them.

  • Fail closed. Unknown stays unknown. A service that is not connected returns NOT_CONNECTED, never success. A saved key reads "Saved, not checked live"; only a recorded successful provider request makes a connection healthy, and only for a day.
  • No fake success. No invented costs, metrics, ads, receipts or connected states.
  • Spending is bounded. Every paid model call is held in a ledger before it is sent, against limits you set on the Billing page. A paid step whose limit is unset is refused, and the refusal says where to set it. A call whose outcome is unknown keeps its hold and is not retried until an administrator releases it, with a reason, in an audited step.
  • Unsafe paths are off. Twenty paid or side-effecting paths stay disabled until they have an idempotency guard and a test: they refuse before any request is made, and they are never retried. The list and the reasons are in docs/OUTBOUND_REQUESTS.md.
  • Outputs are data. Source text, ad text, transcripts, fetched pages and provider outputs are validated and never executed or passed on as instructions.
  • Roles and tenancy. Owners and admins change settings, keys and brands. Members create and review. Viewers read. Every mutation checks the role on the server, and every server function checks the workspace.
  • Secrets stay secret. Keys are encrypted on the server, shown by their last characters only, and never logged.

Run it on your machine

You need Node.js 22. You do not need API keys to start.

npm run dev

The first run installs dependencies, creates .env, and starts the app at http://127.0.0.1:8080. The database is an embedded Postgres file in .data/pglite. Accounts survive a restart. Queued jobs run inside this same process.

Then:

  1. Open the app and choose Create an account instead.
  2. Create a workspace, then a brand.
  3. Follow the Setup guide. The required steps are a brand and a decision engine.
  4. Fill the brand brain. The required fields are positioning, target customers, tone, and prohibited claims.
  5. Add provider keys later, when you have them.
  6. Before any paid step runs, open Billing and set the limits. Nothing is set by default, so paid work is refused until you do.

npm run setup does the .env step on its own and is safe to run again.

Run it on a server

Set DATABASE_URL to PostgreSQL 16. Run the app, npm run worker, and npm run scheduler as separate processes. Production fails closed without DATABASE_URL. Docker Compose is in docs/SETUP.md.

Keys for a workspace are saved in the app. A few values belong to the deployment rather than a workspace, such as the Google Ads developer token. Those are off until you opt in. The names and the switches are in docs/PROVIDERS.md and .env.example.

Where it stands

This is 1.1.0, released as stable by the owner on 2026-11-18. Everything was checked with fixtures, fake transports, a seeded local app, a stand-in for the Claude Code CLI and PostgreSQL in CI. No live provider run is recorded yet. Live provider behavior, real credentials, billing, email and publishing are unverified until the keys are on the server and someone runs them; the live checks still owed are listed in docs/RELEASE_CHECKLIST.md.

What still needs a key or a decision. The code path exists, and it stays not connected until the server has the credential or you set the number:

  • Live image and video calls, including OpenRouter, Deepgram, Drive, and a Hypit process
  • Phrase search through YouTube, Meta Ad Library and SerpAPI. TikTok, Instagram, X, Facebook, Pinterest and LinkedIn are read from public pages only, which often fails; upload the example or paste its text when it does. A site's robots.txt or bot check is reported and never worked around
  • Live publishing and performance for Meta, Google, TikTok, and YouTube
  • Shopify order webhooks, Apify discovery, and invitation email
  • Prices Meridian will not guess: HyperFrames per minute, Apify per 1,000 results, and image.edit

OpenRouter does not state a billing rule for a request that errors or times out, so such an outcome is treated as possibly billed. This is a documented limitation (docs/OPENROUTER.md).

A stored grade is not rewritten when calibration changes. The next run grades the ad again.

Repository layout

Path What is in it
src/routes/ The screens (file-based routing)
src/components/ The interface and the design system in src/components/ui/
src/lib/meridian/ Domain logic and server functions
src/lib/meridian/factory/ The research-led factory pipeline
src/engine/ The video engine. Its specification is docs/engine/VIDEO_ENGINE_BIBLE.md
migrations/ SQL, applied in filename order
scripts/ Setup, the worker and scheduler, and the browser checks

Commands

Command What it does
npm run dev First-run setup, then the app on port 8080
npm run setup Create .env and generate secrets. Safe to repeat
npm run worker Queued jobs, when DATABASE_URL is set
npm run scheduler Periodic work, when DATABASE_URL is set
npm run db:migrate Apply SQL migrations
npm test Unit and integration tests
npm run typecheck Typecheck
npm run lint Lint
npm run build Production build
npm run e2e The product loop in a browser

How it is checked

CI runs the tests in six shards, typecheck, lint and build, the sourcing sidecar's tests, the money checks on a real PostgreSQL server, a browser acceptance of the Settings screens, the UI baseline and accessibility checks, and the product-loop end to end. The check job requires all of them, and it fails if any was skipped. No test calls a live provider. See docs/TESTING.md and docs/RELEASE_ACCEPTANCE.md.

Documentation

Document Read it for
docs/YOUR_FIRST_REAL_VIDEO.md The short, plain-words way to find out whether Meridian makes a video you would post
docs/SETUP.md Install, environment, Drive, deploy
docs/PRODUCT.md The product workflow
docs/FACTORY.md The factory pipeline and the Content page
docs/PROVIDERS.md Every provider and where its key goes
docs/OUTBOUND_REQUESTS.md Every path that can make a request, and how it is guarded
docs/OPERATIONS.md Processes, health, backups
docs/ROADMAP.md What is still open, and why
docs/README.md The full document index
wiki/Home.md The same guide, as wiki pages: setup, the loop, and credentials

About

Advertising for one brand: evidence, the ads, and what happened next. Paid ads are created paused.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages