Inspiration
We drew inspiration from OpenAI’s image processing and multimodal models. Seeing how AI can extract rich insights from visual data sparked the idea to apply similar principles to financial documents—transforming static PDF reports into dynamic, interactive insights.
What it does
FinSightAI ingests your PDF financial reports, automatically parses text, tables, and charts, and indexes the content in a vector database. From there, you can ask natural‑language questions—like “What was our revenue growth last quarter?”—and get clear, data‑backed answers in seconds.
How we built it
- Frontend: A responsive React application styled with Tailwind CSS, featuring a streamlined file‑upload flow and chatbot‑style Q&A interface.
- Backend API: Node.js with Express to handle uploads, serve the frontend, and proxy requests to the Python AI service.
- PDF Ingestion: PDFMiner extracts raw text; we split it into chunks with LangChain’s text splitter.
- Indexing & Retrieval: We store embeddings in FAISS via HuggingFaceEmbeddings for fast similarity search.
- LLM Inference: OllamaLLM powers local, low‑latency question answering using a RetrievalQA chain.
- Data Visualization: Recharts renders key insights and recommended charts in an intuitive dashboard.
Challenges we ran into
- PDF Variability: Handling different layouts, fonts, and table structures required robust pre‑processing and error handling.
- Front‑end Styling: Learning to integrate custom CSS, manage flexbox layouts, and fine‑tune margins within our React components took several iterations to get right.
- API Integration: Securely managing and configuring our OpenAI/Ollama API keys, and handling authentication workflows, presented an initial learning curve.
- UX Polishing: Crafting a clean, intuitive chatbot interface—balancing responsiveness, readability, and visual simplicity—required multiple design refinements.
Accomplishments that we’re proud of
- End‑to‑End RAG Pipeline: A fully integrated Retrieval‑Augmented Generation workflow that reliably processes and answers questions on PDF reports.
- Local Inference: Leveraged Ollama for fast, offline LLM responses without relying on external API calls.
- Engaging UI: Designed a simple yet visually appealing interface that guides users from upload to insights with minimal friction.
- Robust File Handling: Built a production‑ready PDF uploader complete with progress indicators, error recovery, and user feedback.
What we learned
- The importance of data quality—well‑structured text chunks and clean embeddings directly impact answer relevance.
- How to optimize FAISS for real‑time similarity search on modest hardware.
- UI/UX lessons in crafting a conversational interface that feels natural and intuitive.
- Best practices for secure file uploads and sanitizing user input in a public‑facing application.
What’s next for FinSightAI
- Real‑time Alerts & Reports: Push notifications for threshold breaches or emerging trends.
- Better UI Page: Develop a better visually appealing chatbot page with organized charts and chat texts.
- User Profiles & Collaboration: Allow teams to share dashboards, leave annotations, and control access levels.
- Mobile App: A React Native client for on‑the‑go financial insights.
- Expanded Document Types: Support Word, PowerPoint, and HTML reports in addition to PDFs.
- Advanced Analytics: Integrate forecasting models, anomaly detection, and custom KPI tracking.
- Performance Optimization: Improve ingestion and inference speed by batching embeddings, caching frequent queries, and refining model configurations to deliver answers even faster.
Log in or sign up for Devpost to join the conversation.