The challenge
Presentation slides mix text, charts, tables and images, so plain text extraction loses structure and often produces weak summaries or narration.
Multimodal AI
Slide Narrator — Multimodal AI Presentation Intelligence
A multimodal pipeline that understands presentation slides using OCR, table extraction, image captioning and grounded LLM summarization.
Role
Data Scientist / AI Engineer
Context
Selected AI Project
Period
2025–2026
What I built
Designed a slide-aware multimodal pipeline that extracts text with OCR, preserves tables, captions visual elements and stores metadata for each slide element. The LLM then generates grounded summaries and narration with review checkpoints and style adaptation for different audiences.
Architecture
- OCR and layout-aware text extraction
- Table extraction and structure preservation
- Image captioning for visual context
- Slide-level chunking and element metadata
- Grounded summarization with human-in-the-loop review
Outcomes
- Better preservation of slide structure and context
- More grounded narration across text, tables and images
- Audience-aware explanation and summarization workflows
Technology
PythonOCRMultimodal LLMsImage CaptioningRAGStructured Metadata