Your AI Intern That Actually Does the Research

What it does You give it a topic. The backend spins up separate services to draft an outline, search the web (or a local knowledge base), and stream each slide’s content to the page. A Vue frontend lets you edit the outline, pick a template, and watch the pages fill in. The result is a content-complete PPT where data and layout stay separate.

The interesting bit The project treats slide generation as an agentic pipeline—outline agent, slide agent, image search, knowledge-base RAG—rather than a single-shot LLM prompt. It also claims to optimize the agent with GRPO reinforcement learning, though the training code lives in a separate repo, so the exact tuning mechanism is opaque from this codebase alone.

Key highlights

• Separates content from design: pick a template and the agent fills it without mangling the layout.
• Supports both web search and local knowledge-base uploads for sourcing material.
• Modular backend with distinct services for outlines, slide writing, and file parsing.
• Streams content page-by-page to the Vue frontend instead of dumping a finished file.
• Plugs into commercial or local LLMs; you can disable web search entirely by editing agent.py.

Caveats

• The frontend is AGPL-3 because it borrows from PPTist, while the backend claims to be “free copyright”—check licensing before forking for commercial use.
• Custom template upload and auto-annotation are on the roadmap but not yet implemented.

Verdict Worth a look if you need an open, self-hostable alternative to paid AI presentation tools and don’t mind wiring up a multi-service backend. Skip it if you want a polished SaaS experience out of the box.