Introduction

Creating professional presentations is a time-consuming task that requires topic research, structured content writing, visual design, and slide formatting. Automating this process can significantly improve productivity for students, researchers, and professionals.

This project introduces an Agentic AI PowerPoint Builder, a multi-agent system that automatically generates a complete PowerPoint presentation from a user-provided topic. The system uses a coordinated pipeline of AI agents that collaborate to plan slide outlines, perform research, generate slide content, create visual assets, and finally assemble the slides into a .pptx file.

By combining multiple specialized agents with external tools, the system demonstrates how agentic workflows can automate complex creative tasks such as presentation generation.

System Architecture

The system follows a modular multi-agent architecture coordinated by an orchestrator.

Architecture Overview

The execution pipeline is:

1. User Input

2. Planner Agent

3. Research Agent

4. Writer Agent

5. Image Agent

6. Builder Agent

7. Generated PPTX Output

Each agent operates on a shared structured state and updates only the fields it is responsible for.

This separation of concerns ensures:

• Clear role specialization

• Easier debugging

• Better scalability

• Future extensibility

Agents and Their Responsibilities

1) Planner Agent

The Planner Agent analyzes the user topic and creates a structured outline for the presentation.

Responsibilities:

• Understand user topic

• Generate slide titles

• Create slide descriptions

• Define presentation structure

Output:

presentation_outline = [{title, description}]

2) Research Agent

The Research Agent gathers factual information about each slide topic using a web search tool.

Responsibilities:

• Perform web searches for slide topics

• Extract relevant facts

• Generate research notes for each slide

• This improves factual accuracy and reduces hallucination from the language model.

3) Writer Agent

The Writer Agent converts the slide outline and research notes into structured slide content.

Responsibilities:

• Generate bullet points

• Ensure logical flow between slides

• Create concise slide explanations

Output example:

slide_content = [{title, bullet_points}]

4) Image Agent

The Image Agent generates visual assets for each slide.

Responsibilities:

• Generate image prompts

• Retrieve relevant images

• Attach images to slide data

• This enhances the visual quality of the presentation.

5) Builder Agent

The Builder Agent assembles all generated content into a PowerPoint presentation.

Responsibilities:

• Create slides

• Insert text and images

• Format the presentation

• Export .pptx file

The final presentation is then returned to the user for download.

Tools Integrated in the System

The system integrates three external tools, which is a key requirement for agentic workflows.

These tools allow agents to interact with external resources instead of relying only on the language model.

Implementation Details

The system is built using LangGraph to orchestrate agent execution. Each agent is implemented as a node in a state graph and operates on a shared AgentState structure.

Orchestration Flow

The orchestrator manages the following sequence:

1. PlannerAgent → creates presentation outline

2. ResearchAgent → enriches slides with web-based data

3. WriterAgent → generates structured bullet content

4. ImageAgent → generates or fetches visuals

5. BuilderAgent → compiles final PPT

Each agent is modular and independent, improving maintainability and scalability.

The project supports both:

• Streamlit web interface

• CLI-based execution

This dual interface demonstrates flexibility in deployment.

Monitoring and Performance Optimization

To ensure reliability and performance, multiple monitoring mechanisms were implemented.

Structured Logging

A centralized logging system tracks:

• Agent execution stages

• API response times

• Errors and retries

• Final output generation

This allows better debugging and system transparency.

Retry Mechanisms

External API calls use exponential backoff retry logic. If failures occur:

• Web search falls back to LLM-generated content

• Image generation falls back to alternative sources

• Temporary API errors are retried automatically

• This improves resilience in real-world conditions.

Disk Caching

LLM responses and API outputs are cached to:

• Reduce redundant API calls

• Improve response time

• Maintain consistency during testing

• Asynchronous Processing

An optional Redis + RQ job queue enables asynchronous task execution for scalability. If Redis is unavailable, the system automatically falls back to synchronous execution.

Automated Testing

The system includes unit and integration tests using pytest, validating:

• Individual agent correctness

• State transitions

• End-to-end pipeline execution

This ensures system robustness and production-readiness.

Installation

To run the project locally:

git clone https://github.com/dharamshiyash/agentic-ppt-builder
cd agentic-ppt-builder
pip install -r requirements.txt

Create a .env file and configure the required API keys.

Example:

GROQ_API_KEY=your_key
UNSPLASH_ACCESS_KEY=your_key
OPENAI_API_KEY=your_key

The project supports both Streamlit interface and CLI execution.

Usage

Run Web Interface:

streamlit run app.py

Open your browser and enter the presentation topic.

Run Using CLI:

python main.py --topic "Artificial Intelligence in Healthcare"

The system will generate a PowerPoint presentation automatically.

Error Handling and System Resilience

To ensure stable operation, the system implements multiple fallback strategies:

• If web search fails → the Writer Agent uses LLM knowledge.

• If image generation fails → Unsplash images are used as fallback.

• If external API calls fail → retry logic with exponential backoff is applied.

• If slide generation fails → the system logs the error and continues execution.

These mechanisms improve system robustness when interacting with external services.

Applications

The Agentic AI PowerPoint Builder can be used in many real-world scenarios:

-Student presentations

• Research summaries

• Business reports

• Educational content creation

• Automated slide generation for meetings

Key Insights and Learnings

Developing this multi-agent system provided important insights:

• Decomposing complex workflows into agents improves scalability.

• Tool integration significantly enhances output quality.

• Error handling is critical for production AI systems.

• Modular design allows easy extension (e.g., data visualization agents).

• This project demonstrates how agentic AI systems can move beyond simple prompt-based automation toward structured, reliable workflows.

Conclusion

This project demonstrates how multi-agent AI systems can automate complex creative workflows. By combining specialized agents, external tools, and an orchestrated workflow, the system can generate structured presentations from a simple user prompt.

The modular architecture also makes the system easily extendable. Future improvements could include advanced slide design, data visualization, and real-time collaborative editing.