Lovaslide

• 

  Home page.

• 

  AI Agents processing an uploaded document.

• 

  Presentation created from extracted data.

• 

  Successful AI request made to remove the last bullet of slide.

• 

  Successful AI request made to change the theme of the slides from light to dark.

• 

  Another example theme (creative theme) of the presentation.

• 

  Validation checker successfully finds a claim as invalid and provides suggested claims.

Inspiration

Creating presentations is challenging, especially within a constrained time format. So when we looked at other presentation AI copilots, we noticed a common theme: the slides weren't engaging, you have to download them in order to see it, and the tool would be slow and prone to crashes. With Lovaslide (Lovable AI + Slide), We aim to pose a clean and efficient alternative.

What it does

Our platform lets you AI-generate presentations by extracting text from up to three uploaded files — supporting .docx, .pdf, .txt, and .md — and instantly applying one of nine available themes with animated titles, bullet points, and images. Users can preview slides with full functionality and edit presentations effortlessly—adding or removing bullets, pictures, and slides or changing the theme—simply by typing or speaking prompts. The system also grades the accuracy of the information and provides reliable sources to support generated claims, ensuring quality and trustworthiness. Finally, presentations can be exported seamlessly to PowerPoint (.pptx), Google Slides, or PDF, ready for sharing or further refinement.

How we built it

We built Lovaslide as a full-stack AI-powered presentation generator using a modern tech stack. The backend is built with FastAPI in Python, featuring an Analyzer class that uses OpenAI's GPT models to parse uploaded documents (PDFs, Word docs, text files, markdowns) and generate animated, [structured slide content with titles, bullet points, and notes. We implemented a validation agent that extracts factual claims from slides and validates them using both OpenAI for analysis and SerpAPI for web search verification, with built-in rate limiting to handle API constraints gracefully. The backend includes comprehensive document processing using libraries like PyMuPDF for PDFs and python-docx for Word documents, with proper error handling and fallback mechanisms. The frontend is built with Next.js and TypeScript, featuring a sophisticated canvas-based slide renderer that dynamically draws slides with custom themes, animations, and layouts. We created a comprehensive theme system with 9 unique themes each with custom background patterns and color schemes. The frontend includes real-time slide preview, voice note integration using the Web Speech API, interactive validation results display, and export functionality for PowerPoint, PDF, and Google Slides. All these key elements work together to create a seamless AI-powered presentation creation experience.

Challenges we ran into

One of the main challenges we faced was allowing users to upload and process multiple files of different types (PDFs, Word documents, PowerPoints, etc.) at the same time using OpenAI’s API. Each format required different parsing methods, and when combining them, the extracted text often lost structure or contained duplicates. Managing large uploads also created issues with synchronization, performance, and maintaining context across documents. It was difficult to merge overlapping content, preserve headings for slide segmentation, and ensure factual consistency when multiple sources conflicted. Additionally, when one file failed to parse, it could disrupt the entire pipeline, so we had to implement better error handling and workflow reliability to keep the AI slide-generation process stable and accurate.

Accomplishments that we're proud of

One accomplishment we’re especially proud of was successfully integrating voice-to-text functionality to let users interact with the AI agent and modify their presentation slides using speech instead of typing. This feature made the process much more natural and accessible. Users could verbally request edits, add new slides, or refine content hands-free, and the system would interpret and apply those changes in real time. Implementing this required connecting a speech recognition API (Web Speech API), handling background audio processing, and ensuring that spoken commands were accurately converted into structured instructions for the OpenAI-powered agent. It transformed the user experience from a purely text-based interface into an interactive, conversational workflow that felt intuitive and innovative. Another accomplishment we're proud of is the modern-looking UI, which featured each of the agents extracting and parsing the data into slide format and assigning each slide with its appropriate theme, pictures, and animations. which brought that technical look and made the final touch on Lovaslide.

What we learned

The biggest thing that we've learned was how to optimize AI agents for certain tasks, especially when analyzing the text for its most important key points and implementing them through HTML Canvas. Another huge thing we've learned was exposure to different types of APIs—Web Speech API and SerpAPI—and integrating them through FastAPI, all of which played insurmountable roles in our app.

What's next for Lovaslide

We also plan to give the user the option of how many slides to create, dividing the data accordingly. Including an embed of the text where possible and pinpointing the side of validation error would help presenters have more autonomy over their slide management. Additionally, custom fonts and font sizes would be a minor detail that can make the difference between a boring and a more engaging presentation. Finally, being able to upload your own pictures as an optional choice and use Lovaslide to automatically format and preserve them without the output document type.

Built With

• fastapi
• google-web-speech-api
• next.js
• openai
• python
• serpapi
• typescript