_{🐍 Python · 🎬 9:16 MP4 · 🧠 LLM-driven captions · ⚡ 30-60 sec per video · 🤖 End-to-end}

Important

Real production tool. Used at B8 Estate to generate Reels for broker units automatically. Replaces an SMM specialist for content production. See demo_output.mp4 for an example output.

Python pipeline that turns a real estate listing into a TikTok/Reels-style 9:16 MP4 — captions, price/ROI badges, Ken Burns motion. End-to-end automation: broker picks a unit → 30-60 seconds → finished MP4 ready to publish.

🎥 See the result

📺 demo_output.mp4 — click to download/play (3.3 MB, 9:16, 30 sec)

Five generated slides (hero, interior, view, ROI, CTA):

What it does

A marketing engineer for small real estate agencies. Instead of hiring an SMM specialist:

1. Broker clicks "generate video" on a listing in the CRM
2. A Cloud Function calls this Python service with unitId
3. Service pulls photos + project info from Firestore
4. LLM (NVIDIA Llama 70B) generates a 5-frame script with captions
5. Pillow composes 1080×1920 slides with text overlays, badges, brand styling
6. ffmpeg adds Ken Burns motion + music → MP4
7. Result is uploaded back to Storage, broker gets a preview

Replaces the work of an SMM specialist for content production — at scale, on demand.

Architecture

Broker (CRM web app)
  │  POST /api/broker-reel/generate { unitId }
  ▼
Cloud Function (Node.js)
  │  fetches unit + photos + project data
  │  POST → broker_reels HTTP service
  ▼
broker_reels/http_server.py (Flask, on a VPS)
  │
  ├─ firebase_client.py     download photos → /tmp
  ├─ caption_generator.py   NVIDIA Llama → 5-frame script JSON
  ├─ reel_composer.py       PIL slides (photo + caption + price-badge)
  ├─ video_composer (reuse) ffmpeg Ken Burns + music → MP4
  └─ firebase_client.py     upload MP4 to Storage
  │
  ▼  callback POST → Cloud Function `brokerReelCompleted`
  │  writes { reelUrl, reelGeneratedAt } to Firestore
  ▼
Broker sees preview + [Download / Telegram / Regenerate] buttons

Files

File	Purpose
reel_composer.py	PIL 1080×1920 slides with real-estate templates
caption_generator.py	NVIDIA Llama → captions JSON
firebase_client.py	Download photos / upload MP4 via Firebase Admin SDK
orchestrator.py	End-to-end: photos + unit data → MP4
http_server.py	Flask endpoint for the Cloud Function webhook
templates/	Design tokens — colors, fonts, badge templates
demo_photos/	Sample photos for local smoke test
demo_output.mp4	A real generated reel — open to see the output

Five-frame template

#	Frame	Shows	Caption example
1	Hero	Main complex/villa photo	Project name + price badge
2	Interior	Inside shot	"Premium interior · 65 m²"
3	View	Window / location	"5 min from the beach"
4	ROI	Numbers	"ROI 9.1% · 0% installment"
5	CTA	Broker contact	Broker name + handle

Design language

"Warm precision luxury" — navy #003D55 / jade #1E5C5E / gold #A98547 / ivory #FBF7F1. Fonts: Montserrat / Cinzel with safe fallbacks. Animation: subtle Ken Burns (1.0 → 1.05), text fade-in, badge zoom-in.

Local smoke test (no Firebase needed)

cd broker-reels
pip install -r requirements.txt
python3 orchestrator.py --demo   # uses demo_photos/ + hardcoded captions → demo_output.mp4

Tech stack

• Python 3.10+ · Pillow · Flask · requests
• firebase-admin SDK (Storage + Firestore)
• ffmpeg 8.x for video composition
• NVIDIA NIM (Llama 3.3 70B) for caption generation

Deployment

• Local: development & smoke test
• Production: systemd service on a VPS, Flask on port 5000
• Trigger: HTTPS Cloud Function from the CRM

Roadmap

• v1.1: voice clone via ElevenLabs (~$5/mo)
• v1.2: FLUX.1-Depth/Canny to generate missing frames from floor plans
• v1.3: 4 visual presets (morning, sunset, investor, family)
• v1.4: broker watermark + avatar
• v1.5: auto-publish to broker's Telegram channel from CRM
• v1.6: curated cinematic music library (Bensound / Pixabay / Epidemic)

---

⭐ Star history

📚 Related work

• agnis-workflow — the methodology that ships projects like this in 2-3 days instead of 2-3 weeks
• mcp-servers — the marketing/ops MCP toolkit used to publish Reels to Telegram and VK from the broker dashboard
• portfolio — the personal site that ships from this same workflow

---

Author: Sergey Krinitsyn · sergeykrin98@gmail.com · GitHub

_{If this saved you the cost of an SMM specialist — drop a ⭐.}