# Baruch AI — full description

Baruch AI is a self-hosted SaaS that turns products into UGC-style (user-generated-content) video ads. A UGC ad is a short vertical clip in which a relatable creator speaks straight to a phone camera about a product — the native format that performs on TikTok, Instagram Reels, YouTube Shorts and Facebook. Baruch AI produces these without casting, filming or editing.

## Core workflow

### 1. Brand DNA and products
Paste a website URL. Baruch AI fetches the page, extracts the brand's voice, messaging, offerings and content rules, and stores them as a brand kit. Products can be imported from a shop or listing page (title, description, price, photos) or entered manually. Multiple businesses can live in one account, each with its own kit, products, creators, connected accounts, calendar and library.

### 2. Creators
A creator is the on-camera person. Two sources:
- AI creators: a persona (gender, age range, look, outfit, setting, vibe, free-text details) becomes a vertical 9:16 portrait generated with Higgsfield's Soul V2 model and stored in the platform's own storage.
- Uploaded creators: a photo of a real person who has consented to appear in ads.
Creators are reusable and one is marked as the default per business.

### 3. Scripts
The Create wizard asks for: product (or brand video), creator, format, hook style, tone, language, length, number of variants, target platform, optional CTA and notes, plus the video engine and aspect ratio. A text model via OpenRouter (Claude, GPT or Gemini, chosen in Settings) writes 1–5 spoken-word scripts. Each script has 1–4 scenes with a single spoken line and a visual direction, a caption, hashtags and a spoken CTA. Rules enforced in the prompt: hook in the first two seconds, one idea per ad, specific proof over hype, a word budget of about 2.3 words per second, no on-screen text, product facts only from the imported data, and visibly different hooks across variants.

Formats: testimonial, unboxing, problem → solution, POV / day in the life, 3 reasons, before/after, founder story, quick demo.

### 4. Studio
Every script opens in the Studio where the user can edit each scene's dialogue and visual direction, the caption, CTA and hashtags, switch the creator, language, length, aspect ratio and engine, and read the exact prompt that will be sent to the video model. Scripts can be regenerated or rendered from here.

### 5. Rendering
Rendering uses the Higgsfield API. The creator portrait (and product photos, for Seedance) plus the script prompt are submitted to one of:
- Kling 3.0 Standard / Pro — 3–15 seconds, vertical (follows the creator portrait), native lip-synced speech
- Kling 2.6 Pro — 5 or 10 seconds, 9:16 / 1:1 / 16:9, native speech
- Seedance 2.5 — 4–30 seconds, 9:16 / 1:1 / 16:9, reference-to-video with creator and product images together, native speech
Spoken lines are wrapped in quotes in the prompt so the model generates the voice and lip movement with the video. Renders run in a persistent background worker that polls the provider (and accepts completion webhooks when the server is public), so closing the browser never loses a video. Finished MP4s are copied to the platform's own storage because provider URLs expire after about seven days. Credits are reserved when a render starts and refunded automatically if it fails, is rejected by content moderation, or times out.

### 6. Library and publishing
The library shows every ad as a 9:16 card with its status (script ready, queued, rendering, ready, failed), format, engine and creator. Ready videos can be downloaded or published. Publishing targets connected accounts:
- Instagram Business / Creator — Reels (video_url container → publish)
- Facebook Pages — Page video
- LinkedIn — video post (initialize upload → upload → finalize → post)
TikTok and YouTube Shorts are tracked with a one-click "Mark as published" flow. Demo installs can turn on a simulation mode that produces fake URLs without posting.

### 7. Calendar
Slots on the content calendar hold a brief, creator, format, engine and length. About 15 minutes before a slot is due, the worker writes the script and renders the video so the file is ready at publish time; the slot shows "Rendering video" until then and publishes automatically when the MP4 lands. "Plan with AI" fills a date range with varied UGC ad angles across products, formats and platforms.

## Credits and pricing
One credit balance is shared by everything:
- Script writing: the text model's tier cost (Light 1, Standard 2, Pro 4, Premium 10) per variant
- AI creator portrait: flat (default 1 credit)
- Video render: clip seconds × the engine's credits-per-second rate, which the platform administrator sets per engine against Higgsfield's list price
Credit packs are one-time purchases through Stripe, PayPal, Razorpay or Dodo Payments; credits never expire. Free-trial accounts get starter credits, Light-tier text models and the budget video engine; any purchase unlocks all models for life.

## Administration
Admins manage users and credits, credit packs, payment gateways, the OpenRouter key, the Higgsfield API key (with a connection test), per-engine credit pricing and enablement, trial defaults, demo/simulation mode, and a gallery of every rendered ad with render health and provider usage.

## Technology
Next.js (App Router) and React frontend; Node.js, Express and MongoDB backend; OpenRouter for LLMs; Higgsfield API for video (Kling, Seedance) and portraits (Soul); UploadThing for storage; Brevo for transactional email; Meta Graph API and LinkedIn REST API for publishing. Long-running jobs stream progress over server-sent events and are tracked in a persistent task list.