B-roll Assistant
Project: March 2025 — November 2025
Role: AI / Backend Engineer
Client: PAP (Polish Press Agency)
Project description
B-roll Assistant is an intelligent AI assistant that automatically catalogs video materials and enables their instant search.
In newsrooms and video production environments, time is critical. Editors and video editors must quickly find the right B-roll, often under deadline pressure. In practice:
- video materials are poorly described or not at all,
- file names are ambiguous,
- manual tagging is costly and unrealistic,
- classic folder-based search doesn't scale.
The system:
- analyzes video without user intervention,
- generates snapshots and shot descriptions using AI,
- combines classic search with semantic search,
- learns from editor ratings.
Thanks to this, the editor doesn't need to know how the file was named — just describe what they're looking for.
How It Works
1. Smart Catalog Setup
- System adapts to existing file structure
- No manual organization required
- Handles large B-roll archives
2. AI Shot Descriptions
- Automatic snapshot generation (FFmpeg)
- Visual shot analysis (AI Vision)
- Text descriptions for each clip: objects, scenes, context, atmosphere
- Each clip receives a rich semantic description without manual tagging
3. Dual Search & Continuous Learning
- Keyword search (fast, precise queries)
- Semantic search (search by meaning and concepts)
Example queries:
- "glass"
- "transparent container"
- "tense political rally"
- "crowd waiting nervously"
The system enables rating results, which:
- improves accuracy,
- adjusts ranking,
- increases search effectiveness over time.
Architecture Overview
B-roll Assistant is a multimodal pipeline:
- Video ingestion — file structure analysis, frame extraction
- AI processing — vision → image description, text → embeddings
- Search layer — classic indexes, vector database
- Feedback loop — user ratings, result ranking correction
Designed for performance and scaling in media environments.
What I did
- Designed the B-roll analysis system architecture
- Implemented video frame extraction pipeline
- Integrated AI Vision for shot analysis
- Created automatic clip description system
- Built dual search engine (keyword + semantic)
- Deployed embedding-based search
- Designed user feedback mechanism
- Prepared backend API and data structure
- Containerized the system (Docker)
Skills
- Python
- OpenAI (Vision + LLM)
- Embeddings
- Vector Database
- FastAPI
- PostgreSQL
- FFmpeg
- Docker
- Semantic Search
Results
- Find the right B-roll in seconds instead of minutes
- No need for manual tagging
- Better utilization of existing video archives
- System that learns along with the editorial team
- Solid foundation for: newsrooms, media houses, content platforms
Sample photos
