Why Marée?
The modern marketing and design landscape is fragmented. Teams constantly switch between disjointed platforms, copy-pasting sensitive data across unsecure cloud services while trying to navigate rapidly shifting compliance and data-privacy regulations.
Marée brings your entire workflow under one roof. From company profile ingestion to guided campaign strategies, everything stays local and under your direct control. Whether you want to operate in absolute "stealth mode" entirely offline or selectively integrate cloud APIs and connectors (such as Google Drive, Obsidian, or Figma), Marée adapts to your security requirements—not the other way around.
What It Is
Marée is a modular agency platform built for marketing and communications teams that want full control over their AI infrastructure. It covers the entire agency workflow: from market research and campaign strategy, through content creation and brand design, to scheduled publishing and performance analysis.
The platform runs locally on your hardware. All AI features are powered by on-premise large language models (Gemma-4-12b qat-Q4) via CUDA — no external API keys required, no subscription fees, no training on your data.
- SAGE → Market intelligence & topic discovery
- Breeze → Campaign strategy, content planning, post production, AI campaign review, performance forecast & recommendations
- SeaDesign → Brand identity, design system & visual content creation
- Leaflet Still → AI-powered still image generation
- Cron Publisher → Automated, scheduled multi-platform delivery
Breeze — Full Campaign and Forecast Suite
The core campaign management workspace. Covers the full lifecycle from strategy brief to post review.
Campaign Intelligence
- Upload market research, competitor analyses, TAM/SAM/SOM documents, and customer interviews
- Map-Reduce engine processes any number of documents — each is summarized individually, then synthesized into a structured strategy brief
- Extracts: ICP, audience pains, positioning, messaging angles, competitive whitespace, market size, recommended campaign objective
Campaign Planner
- Define campaign parameters: name, platform, ICP, seniority level, target region, objective, budget, start date
- Automatic forecast projection: leads, impressions, and click-through rates over the campaign duration — with confidence bands (low/high range) and status color coding
- Campaign phase visualization (Awareness / Consideration / Conversion) with week-by-week breakdown and German holiday markers
- Content-type timeline: color-coded circles per post (Video, Explainer, Ad, Case Study, Thought Leader, etc.)
Post Production
- Structured post editor: headline, body, CTA, psychological hook, target audience, format type
- Slide-level management: upload images per post slide, replace or add slides
- Review workflow: OK / Review / Incomplete status per post
- Posting calendar with platform-specific best-time recommendations
- Campaign restructure analyst: AI-driven full campaign replanning with a checkbox-modal for selective application
Campaign Analyst (AI Chatbot)
- Powered by Gemma-4-12b qat-Q4 — deterministic Python pre-analysis runs before LLM call: post completeness flags, CTA gaps, headline/body length checks, content mix distribution, psychological hook coverage
- Chat interface with suggestion chips, conversation history, and inline post references
- Frameworks applied: AIDA, Pain-Gain, Message-Market-Fit, Platform-Native (LinkedIn algorithm requirements)
- Execute-intent detection: "apply all suggestions" triggers restructure plan directly in chat
SeaDesign — Chat-based Design System & Editing Suite
Design system creation, visual editing, and AI-assisted brand generation with multi-format export.
Design Systems
- Create, version, and manage multiple brand design systems in parallel
- Each system stores: color palette, typography, spacing tokens, component styles, logo SVGs, and reference images
- CSS variable-based theming — systems apply brand colors and fonts directly to AI-generated designs
- Folder/file ingest: drop in brand assets and the platform extracts tokens automatically
AI-Assisted Design Generation
- Describe a design in natural language — the agentic loop (create / view / str_replace / done) generates and iteratively refines a standalone HTML artefact
- Format presets: LinkedIn Post (1200×627), 1:1, Portrait (1080×1350), Banner, Story/Reel, Instagram, Logo SVG
- Polish mode: one-click design refinement for hover states, realistic content, and visual detail
- Template library: auto-generated from design system assets
Visual Editor
- Click any element in the live preview to select it — overlay highlights with element info
- Properties panel: Layout, Size, Position, Spacing (margin/padding box), Typography, Background, Border, Effects, Transform, SVG
- 9-tile alignment grid: set
align-self,justify-self, and margin-based positioning in one click - Lock elements (
data-ceed-locked): the LLM instruction layer and server-side enforcement prevent locked elements from being modified by any AI edit
- Target elements (
data-ceed-target): mark an element as the chat target — prompts like "center the logo" auto-inject[Ziel: logo] - Replace panel: swap text, fonts, images, or SVG assets directly from the design system
- Copy element to other slides with one click
- Layers panel: tree view of all elements in the current slide
- Undo stack for all style changes
Leaflet Still
Image Generation- Generate images from text prompts via local Flux Schnell (CUDA-accelerated)
- Designed for: post hero images, background visuals, product shots, abstract brand graphics
- Output formats: PNG at standard social media resolutions
- No cloud API — fully local inference on RTX 3090
Dashi
Weekly Market Digest- Aggregates relevant topics from configured sources
- Clean, readable layout in light and dark mode
- Connects topic signals to active Breeze campaigns
- Export-ready for distribution to clients or teams
Cron Publisher
Automated Scheduling- Cron syntax scheduling: post to platform, send digest, trigger analysis
- Markdown publisher: converts campaign content to formatted Markdown for automated delivery
- Pluggable output targets: file export, webhook, platform API
- Research hooks: trigger scheduled intelligence runs to keep strategy briefs current
LLM Infrastructure & Architecture
Marée is built on a local, CUDA-accelerated model stack. All AI features run on your hardware.
| Model | Role | VRAM |
|---|---|---|
| Gemma-4-12b qat-Q4 | Campaign analysis, strategy synthesis, complex reasoning, design generation | ~8 GB (4-bit Q4_K_M) |
| GLM-OCR | Document scanning, image-to-text extraction | ~2 GB |
Requirements: NVIDIA GPU with CUDA support (tested on RTX 3090 24 GB), Python 3.11+, PyTorch 2.6 + CUDA 12.4
Environment: gemma_venv. All calls run with allow_external=False by default — the platform will not route requests to any cloud API unless explicitly configured per-session.
Privacy guarantee: Documents, campaign briefs, post content, and brand assets never leave your network. No telemetry. No external model calls in default configuration.
Architecture Stack
Python 3.11 · Flask · SQLite · React 18 · Vite · Lucide Icons · ReportLab · Playwright
Designed For
- Marketing agencies running B2B LinkedIn campaigns for multiple clients
- In-house marketing teams that need campaign intelligence without exposing data to cloud AI
- Consultancies building strategy-first content programs at scale
- Teams with GPU capacity who want to run professional AI tooling without recurring API costs
Ready to claim sovereignty over your agency workflow?
Marée is a proprietary system — all rights reserved. Built with PyTorch 2.6, CUDA 12.4, React 18, Gemma-4-12b qat-Q4, Playwright, and ReportLab.
Contact for Licensing Inquiries