Auto_PPT_Agent

> An AI agent that autonomously researches, plans, and generates PowerPoint presentations from a single sentence prompt - powered by a custom MCP (Model Context Protocol) architecture.

[Watch Demo Video (2 min)](https://drive.google.com/file/d/1LMYYZzZ_BmHVUnYhxRYoEmX114bAZrPF/view?usp=sharing)

---

Auto-PPT Agent is a full-stack application built for the AI Agents & MCP Architecture course. Given a prompt like *"Create a 10-slide presentation on Quantum Computing"*, the agent:

1. Plans the slide outline using an LLM (Groq / HuggingFace)
2. Researches each slide topic via DuckDuckGo / Wikipedia
3. Generates a relevant AI image per slide (HuggingFace / Pollinations)
4. Builds the .pptx file slide-by-slide through MCP tool calls
5. Streams real-time progress to a React frontend over WebSocket

All of this happens without the user doing anything after clicking Generate Deck.

---

1. Code Repository - Complete source code for the agent, backend, frontend, and all MCP servers included. 2. Video Demo - Link to 2-minute demo video showing the agent creating a PPT from scratch. 3. Reflection Document - Attached as MekalaSanjith_Auto_Agent_PPT_Reflection.pdf, answering the prompt about where the agent first failed and how MCP prevented hardcoded logic.

---

The agent follows a deterministic agentic loop - it must explicitly plan before it can execute. This is enforced by the submit_slide_plan gate in the PPTX MCP server.

User Prompt
    │
    ▼
FastAPI Backend (WebSocket)
    │
    ├─► LLM (Groq / HuggingFace)
    │       └─► Plan N slide titles
    │
    └─► MCP Agentic Loop
            │
            ├─► MCP #1: pptx_mcp_server
            │       create_presentation → submit_slide_plan (gate)
            │       add_slide_with_image / add_slide (per slide)