AI Document Intelligence Platform

This is the build — we called it Catalyst — that won SwissHacks 2024 outright: a double win, taking both the Unique/Microsoft Challenge and the Overall Hackathon out of a field competing over 48 hours. The brief: turn a dense, unstructured financial PDF into decisions, fast enough that there is no waiting between uploading a document and interrogating it.

The hard part: zero-latency, not just "an LLM"#

Anyone can wire a PDF to an LLM. The reason this won was that the judges could upload a document and get answers with no perceptible wait — no spinner while a model chewed through the whole file. Getting there meant building a real retrieval pipeline rather than stuffing a document into a prompt.

The RAG architecture#

The key decision, and the one the overall win hinged on, was to combine proven open-source building blocks and cloud infrastructure so we could move at hackathon speed without hand-rolling the hard parts:

1. High-speed parsing. An ingestion pipeline analyses the uploaded PDF in real time, structures it, and surfaces the most important statistical key points immediately — so the first useful view paints before the document has finished processing end to end.
2. Vector search & embeddings. We used Pinecone to index and store the document's embeddings, which is what made semantic search feel instant: retrieval is a nearest-neighbour lookup, so pulling the right context for a question does not scale with document size.
3. Grounded generation. Only the retrieved, relevant chunks go to the OpenAI model, so answers are grounded in the document and cheap to produce rather than re-reading the whole file per question.

The interface that sold it#

We wrapped it in a three-column dashboard that matches how someone actually reads a financial document under time pressure:

• Left — aggregated KPI metrics and decision-making keywords, for the ten-second overview.
• Middle — the document itself.
• Right — an interactive AI chat pinned to the document, for the follow-up questions the metrics raise.

Why it was intense#

Two winning submissions in one weekend meant no slack anywhere: the parsing had to be fast, the retrieval had to be accurate, and the UI had to make all of it feel effortless — built in continuous night shifts as a team, holding one clear vision without letting performance slip.

Stack: OpenAI API · Pinecone (vector database) · Python · FastAPI · Next.js · Vercel.