Ad-Genie

AI-powered ad creation platform that generates images, videos, audio, and copy from a single brief.

Comment

Ad Genie: AI-Powered Ad Creation Platform

Inspiration

The inspiration for Ad Genie came from a simple observation: creating high-quality advertisements requires a diverse skill set — copywriting, visual design, video production, audio engineering, and platform-specific optimization. Small businesses and solo creators often lack the budget to hire agencies or the time to master multiple tools. We asked: What if a single AI system could generate a complete, platform-ready ad campaign from just a text brief?

The breakthrough came with the emergence of Pollinations.ai — a free, no-API-key-required generative AI service offering image (Flux), video (Minimax/Hailuo), and audio (TTS) generation. This democratized access to multimodal generation and became our primary provider, with premium fallbacks (OpenAI Sora, ElevenLabs, GMI Cloud) for users needing higher fidelity.

What it does

Ad Genie is an end-to-end AI ad generation pipeline that transforms a natural-language brief into production-ready assets:

Asset Type Primary Provider Fallback Providers Output
Images Pollinations (Flux) GMI Cloud (Seedream/Flux), OpenAI (gpt-image-1, DALL·E 3) 1024×1024 – 1920×1080
Video Pollinations (Minimax-video-01-live) GMI Cloud (Seedance, Kling), OpenAI (Sora) 1080×1920, 4–12s
Audio Pollinations (TTS) ElevenLabs (v3), OpenAI (tts-1-hd), Stability (Stable Audio) MP3/WAV, voiceover + music
Copy OpenRouter (DeepSeek V4) OpenAI (GPT-4o) Headlines, body, CTAs, scripts

Key features:

  • 🎯 Campaign workflow: Brief → Business profile → Asset variants → Library
  • 🔐 Auth + per-user isolation: Supabase Auth + Row-Level Security
  • ☁️ Durable storage: Backblaze B2 (S3-compatible) with signed URLs
  • ⚡ Async generation: Genblaze Python service (FastAPI + Uvicorn) handles long-running video/audio jobs
  • 🎨 Platform presets: TikTok (9:16), Reels (9:16), YouTube (16:9), Feed (1:1, 4:5)
  • 📊 Analytics-ready: SHA-256 hashes, latency tracking, retry logic

How we built it

Architecture

graph LR
    A[User Brief] --> B[TanStack Start Frontend]
    B --> C[Supabase Auth & DB]
    B --> D[Server Functions]
    D --> E[Genblaze Service :8000]
    E --> F[Pollinations.ai]
    E --> G[OpenAI / GMI / ElevenLabs]
    E --> H[Backblaze B2]
    H --> I[Signed URLs → Frontend]

Tech Stack

Layer Technology
Frontend TanStack Start (React 19, Vite, SSR), Tailwind CSS 4, Radix UI
Backend TanStack Start server functions, Supabase (PostgreSQL + Auth + RLS)
Generation Python Genblaze service (FastAPI, Uvicorn, httpx, pydantic)
Storage Backblaze B2 via aws4fetch (SigV4 signing)
AI Providers Pollinations.ai (primary), OpenAI, GMI Cloud, ElevenLabs, Stability, OpenRouter
Infra Netlify (frontend), Docker/VM (Genblaze), Supabase Cloud

Core Components

1. Provider Abstraction (src/lib/adforge.functions.ts)

const GENBLAZE_PROVIDERS = {
  image: [
    { provider: "pollinations", model: "flux", free: true },
    { provider: "gmicloud", model: "seedream-5.0-lite" },
    { provider: "openai", model: "gpt-image-1" },
  ],
  video: [
    { provider: "pollinations", model: "minimax-video-01-live", free: true },
    { provider: "gmicloud", model: "seedance-2-0-260128" },
    { provider: "openai", model: "sora-2" },
  ],
  audio: [
    { provider: "pollinations", model: "tts-1", tier: "standard", free: true },
    { provider: "elevenlabs", model: "eleven_v3", tier: "premium" },
    { provider: "openai", model: "tts-1-hd", tier: "professional" },
  ],
};

2. Pollinations Direct Path (bypasses Genblaze for speed)

if (provider === "pollinations") {
  if (assetType === "video") {
    const url = `https://gen.pollinations.ai/video/${encodeURIComponent(prompt)}?model=minimax-video-01-live&duration=${durationSec}`;
    // fetch → B2 PUT → return signed key
  }
  // ... image & audio similar
}

3. B2 Storage with SigV4 (src/lib/b2.server.ts)

function client() {
  return new AwsClient({
    accessKeyId: env("B2_KEY_ID"),
    secretAccessKey: env("B2_APPLICATION_KEY"),
    region: env("B2_REGION"),  // eu-central-1 (AWS-compatible for signing)
    service: "s3",
  });
}

4. Genblaze Service (python-genblaze/genblaze_service.py)

class GenerationRequest(BaseModel):
    prompt: str
    provider: str
    model: str
    asset_type: Literal["image", "video", "audio"]
    width: int = 1024
    height: int = 1024
    duration: Optional[int] = None
    # ...

@app.post("/generate/async")
async def generate_async(req: GenerationRequest):
    run_id = str(uuid4())
    background_tasks.add_task(execute_generation, req, run_id)
    return {"run_id": run_id, "status": "pending"}

@app.get("/generate/status/{run_id}")
async def generation_status(run_id: str):
    # returns { status, assets: [{url, mime_type, sha256}] }


## Challenges we ran into

### 1. **OpenAI Billing Hard Limits** 💳
> `Error code: 400 - Billing hard limit has been reached`

OpenAI's `gpt-image-1` and Sora require prepaid credits with strict limits. This forced us to make **Pollinations the default** and build intelligent fallback chains.

### 2. **Genblaze `file://` URI Handling on Windows** 🪟
Python's `Path.resolve()` on `file:///C:/...` URIs fails. Fixed in `genblaze_core/storage/transfer.py`:
```python
# Before (broken on Windows)
Path(unquote(parsed.path)).resolve()

# After (correct)
from urllib.parse import url2pathname
Path(url2pathname(unquote(parsed.path))).resolve()

3. Sora SDK Videos Object Bug 🐛

OpenAI's Python SDK returns a Videos object without .content. Patched genblaze_openai/provider.py:

# Download via HTTP instead of relying on .content
async with httpx.AsyncClient() as client:
    resp = await client.get(video.url)
    resp.raise_for_status()
    content = resp.content

4. B2 Region Validation 🌍

aws4fetch rejects eu-central-003 (B2-specific). Solution: use AWS-compatible region eu-central-1 for signing only; actual endpoint comes from B2_ENDPOINT.

5. Supabase Project Mismatch 🔄

.env pointed to empty project shnmfnweneysrsxrdcdm while real project is nxkjjwxlgnizqsrvegba. Caused blank images (RLS blocked inserts). Fixed by aligning env vars.

6. Dual Python Site-Packages 📦

Genblaze installed in both venv/ and WindowsApps global site-packages. Both must be patched for fixes to take effect.

Accomplishments that we're proud of

Achievement Impact
Pollinations as universal free tier Zero-cost image/video/audio for all users
End-to-end video pipeline Sora job eaad1ca9 → 12.8 MB MP4 → B2 → signed URL ✅
Unified provider abstraction Single API for 6+ AI providers with automatic fallback
Windows-compatible Genblaze Fixed file://, Sora SDK, region signing — runs natively
Per-user asset isolation RLS + signed URLs = zero data leakage
Async generation with polling Long video jobs don't block UI; status endpoint + WebSocket-ready
SHA-256 integrity + latency tracking Every asset verifiable; performance measurable

What we learned

  1. Free tiers unlock adoption — Pollinations removed the biggest barrier (API keys + billing). Users try the product instantly.

  2. Provider diversity = resilience — No single provider is 100% available. Fallback chains with automatic retry are essential.

  3. Windows Python paths are treacherous — file:// URIs, dual site-packages, and WindowsApps Python require explicit handling.

  4. Signed URLs > public buckets — B2 + aws4fetch SigV4 gives secure, expiring access without CORS headaches.

  5. Async first, sync later — Video generation takes 30–180s. Designing for async from day one prevented major refactors.

  6. Environment parity matters — .env must match across frontend, server functions, and Genblaze service.

What's next for Ad Genie

Near Term (v1.1)

  • [ ] Webhook callbacks from Genblaze → frontend (replace polling)
  • [ ] Asset variants UI — A/B test multiple generations per brief
  • [ ] Brand kit — Logo, colors, fonts injected into prompts automatically
  • [ ] Cost tracking — Per-provider, per-campaign spend dashboard

Medium Term (v1.5)

  • [ ] Video understanding — Analyze generated videos for brand safety / quality scoring
  • [ ] Multi-language TTS — Pollinations voices + ElevenLabs multilingual
  • [ ] Template marketplace — Community-contributed prompt templates
  • [ ] Direct social posting — Meta, TikTok, YouTube API integration

Long Term (v2.0)

  • [ ] Agentic campaign optimizer — Auto-iterate based on performance data
  • [ ] Custom model fine-tuning — LoRA on brand assets for consistent style
  • [ ] Real-time collaboration — Multiplayer editing with conflict resolution
  • [ ] Enterprise SSO + audit logs — SAML/OIDC, SOC2 compliance

Built With

  • aws4fetch
  • backblaze-b2
  • deepseek
  • docker
  • elevenlabs
  • fastapi
  • generative-ai
  • gmi-cloud
  • netlify
  • openai
  • openrouter
  • pollinations-ai
  • postgresl
  • python
  • react-19
  • sora
  • stability-ai
  • supabase
  • tailwindcss-4
  • tanstack-start
  • typescript
  • uvicorn
  • vite
Share this project:

Updates