This is a submission for the Built with Google Gemini: Writing Challenge
What I Built with Google Gemini
As a computer science student working on research projects, I often face the same problem: research papers are difficult and time-consuming to understand. Papers can be long, technical, and full of complex equations, making it hard to quickly extract the key ideas.
To solve this, I built PaperLens AI, an application that allows users to upload a research paper PDF and interact with it using AI.
The app automatically:
- Extracts text from the uploaded research paper
- Splits the document into semantic chunks
- Creates embeddings for each chunk using Google Gemini embeddings
- Retrieves the most relevant sections of the paper
- Uses Gemini to answer questions about the paper
In simple terms, PaperLens AI lets users chat with research papers.
For example, users can ask questions like:
- What is the research gap in this paper?
- What dataset was used?
- What are the limitations of the method?
The AI then finds the most relevant sections of the paper and generates an answer based on those sections.
Architecture of the System
The system uses a simple Retrieval Augmented Generation (RAG) pipeline.
Research Paper PDF
↓
Text Extraction (PyPDF2)
↓
Chunking
↓
Gemini Embeddings
↓
Similarity Search
↓
Gemini generates answer
Technologies used:
- Python
- Streamlit (UI)
- Google Gemini API
- NumPy / Scikit-learn
- Vector similarity search
The goal was to create something lightweight but powerful enough to help students quickly understand complex research.
Demo
Here is the interface of PaperLens AI.
aggreypaintsil168
/
paperlens-ai
Research papers are difficult and time-consuming to understand. Papers can be long, technical, and full of complex equations, making it hard to quickly extract the key ideas.
📄 PaperLens AI
Chat With Research Papers Using Google Gemini
🚀 Overview
PaperLens AI is an AI-powered application that allows users to upload research papers and interact with them conversationally.
Instead of reading long academic papers line by line, users can ask questions like:
What is the research gap in this paper?
What dataset was used?
What methodology does the paper propose?
What are the limitations of the study?
PaperLens AI retrieves the most relevant sections of the paper and uses Google Gemini to generate contextual answers.
This transforms static research documents into an interactive knowledge system.
🎥 Demo
Example UI:
Upload Research Paper
↓
Extract + Chunk Paper
↓
Create Embeddings
↓
Ask Gemini
Example query:
What is the research gap of this paper?
Example output:
The research gap addressed by this paper is the inability of
traditional forecasting models to capture both temporal
dependencies and trendUsers can:
- Upload a research paper
- Extract and chunk the document
- Generate embeddings
- Ask questions about the paper
Example workflow:
Upload PDF
→ Extract + Chunk Paper
→ Create Embeddings
→ Ask Gemini
Example question:
What is the research gap in this paper?
The system retrieves relevant sections and Gemini generates an explanation.
(If deployed, the Cloud Run app would be embedded here.)
What I Learned
This project taught me several technical and conceptual lessons.
1. How Retrieval-Augmented Generation Works
Before this project, I had only read about RAG systems. Implementing it myself helped me understand how modern AI assistants work internally.
Instead of sending an entire document to the model, we:
- Convert text into embeddings
- Store them as vectors
- Retrieve the most relevant pieces of text
- Send those pieces to the AI model
This dramatically improves both accuracy and efficiency.
2. Building AI Applications Requires More Than Just Calling an API
The real engineering work was in:
- Text chunking
- Embedding generation
- Similarity search
- Prompt engineering
These components are what transform a basic model call into a useful application.
3. Debugging AI Infrastructure
While building PaperLens AI I encountered several issues, including:
- API authentication errors
- Deprecated Gemini models
- Embedding model changes
- API quota limitations
Debugging these issues taught me a lot about how real AI systems evolve quickly, and why developers must constantly adapt to new APIs and model updates.
4. Building AI Tools for Research
As someone interested in machine learning and research, building this project made me realize how powerful AI can be for assisting with scientific work.
Tools like this could help students:
- Understand research papers faster
- Identify research gaps
- Generate ideas for new research directions
Google Gemini Feedback
Overall, Google Gemini made it relatively easy to build a powerful AI application.
What worked well
The embedding capabilities and text generation features made it possible to build a full RAG pipeline with relatively little code.
Gemini was particularly good at:
- Explaining complex technical concepts
- Summarizing research sections
- Answering contextual questions about documents
Where I ran into friction
One challenge I encountered was related to model changes and deprecations.
During development, some models referenced in earlier documentation were no longer available, which required updating the application to use newer Gemini models.
Another challenge involved API quota limits, which temporarily prevented some requests from completing during testing.
These issues highlighted how quickly AI platforms evolve and the importance of keeping projects updated with the latest SDKs and models.
What I’d like to see improved
Two improvements that would make Gemini development even better are:
- clearer versioning for model releases
- easier visibility into quota usage during development
These would help developers avoid unexpected errors while building applications.
What’s Next
PaperLens AI is just the beginning.
Future improvements could include:
- Multi-paper comparison
- Automatic research gap detection
- Citation extraction
- Integration with academic databases
Ultimately, I’d love to turn PaperLens AI into a tool that helps students and researchers navigate the growing volume of scientific literature.
Final Thoughts
This project reminded me that the most exciting part of AI development isn’t just the models themselves — it’s how we apply them to real problems.
For students and researchers, understanding complex papers is often a major barrier. With tools like Google Gemini, we can start building systems that make knowledge more accessible.
And that’s something worth building.
Top comments (0)