I built a real AI video processing SaaS from Senegal no GPT wrappers, just HuggingFace + OpenCV + YOLO + Detectron2+Medidapie+ Celery

Source: Dev.to

Problem I was solving

Every creator I know spends 3–4 hours manually cutting video.
The algorithm rewards volume — not perfection.

Solution Overview

I built ClipFarmer, a SaaS that processes video entirely with on‑premise machine‑learning models instead of third‑party API wrappers.

Machine Learning Models

• Whisper (HuggingFace) – automatic speech transcription.
• YOLO + OpenCV (cv2) – scene detection.
• Detectron2 – instance segmentation.
• MediaPipe – pose and face landmark detection.
• OpenCV (cv2) – backbone for all frame‑level operations.

These are real models running locally; no external API calls.

Effects Pipeline

Each effect is a cv2 pipeline that processes frames:

• Color grading (dark moody, vintage grain, RGB split)
• CRT scanline overlay
• Motion blur
• Skeleton overlay (MediaPipe pose)
• Background removal (Detectron2 masks)
• Transitions between clips using frame blending

Architecture

Backend

• FastAPI + Celery + RabbitMQ + Redis

AI / Computer Vision Stack

• Whisper, YOLO, Detectron2, MediaPipe, OpenCV

Storage

• MinIO (self‑hosted S3‑compatible, presigned uploads)

Frontend

• React + Vite + TailwindCSS

Database

• PostgreSQL + SQLAlchemy (async)

Deployment

• Docker Compose on a VPS

Task Orchestration

# Celery chord that runs the whole pipeline
workflow = chord(
    spliter_clip.s(job.job_id, input_path),
    workflow_tasks_parallel.s()
)
task_result = workflow()

The workflow first splits the video, then applies effects, subtitles, and transitions in parallel.

Regional Considerations (Senegal & West Africa)

• Mobile money (Wave, Orange Money) is the primary payment method; credit cards are rare.
• ClipFarmer accepts Wave and Orange Money natively.
• Many AI tools seen locally are scams or inaccessible, so providing a locally hosted solution is crucial.

Challenges Faced

• Conflicting ML dependencies across models.
• Presigned uploads are mandatory for large video files.
• cv2 frame processing is slow without proper batching.
• Docker networking can be unexpectedly restrictive.

Availability

Live at clipfarmer.site – free credits are available for testing.

Call for Feedback

I’m curious: has anyone else built a cv2‑based processing pipeline? What would make you switch from manual editing to an automated solution like ClipFarmer?