At HackMIT, I built Nest in the YC track, which was a challenge to take a pre-ChatGPT YC company and reimagine it for the AI era. I chose Dropbox.
Dropbox is great for storing files, but it was built for people, not AI. You upload a research paper, a slide deck, some meeting notes, and they just sit there. No context. No relationships. No understanding of how they connect.
Nest is an AI-native document platform. It doesn't just store files—it gives Claude direct access to your team's knowledge through smart summaries, a 3D knowledge graph, and an AI agent that actually understands what's in your docs. This agent can also handle tasks for you, such as sending documents or receiving signatures.
The Knowledge Graph
The centerpiece is a 3D visualization that maps all your documents as nodes in space. I used a Fibonacci sphere algorithm to distribute them uniformly, and the graph automatically scales based on how many documents you have. But here's the cool part: when you click on a connection between two documents, Claude streams a real-time explanation of how they relate. This dynamically changes as more documents are added in context.
How It Works
The pipeline starts when you drag in a PDF or document. The FastAPI backend extracts the text using PyPDF2, then splits it into semantic chunks: 512 tokens each with 50-token overlap, chunked at sentence boundaries using NLTK to preserve meaning. Each chunk gets embedded with OpenAI's text-embedding-3-small model and stored in ChromaDB.
While that's happening, Claude analyzes the full document to generate a concise summary and five relevant tags. The frontend updates in real-time via Supabase subscriptions, so if your teammate uploads something, you see it immediately.
There's also a chat interface where you can ask questions about your documents. The AI agent queries ChromaDB for the most relevant chunks (similarity threshold > 0.3), passes them to Claude with your question, and streams back an answer with source citations. Beyond Q&A, the agent can also complete tasks for you—sending documents, requesting signatures, or sharing files with teammates.
Scaling with AWS Kinesis
A document platform is useless if it can't handle high concurrent load. We integrated AWS Kinesis for data streaming to ensure the system could handle thousands of concurrent uploads without bottlenecks. When users upload documents simultaneously, Kinesis buffers and streams the data to our processing pipeline, preventing any single component from getting overwhelmed. The entire infrastructure is hosted on AWS and built to scale horizontally, so whether it's 10 users or 10,000, the system stays responsive.
Tech Stack
Next.js 15 for the frontend, FastAPI + Python for the backend. Claude handles all the AI heavy lifting (summaries, tags, Q&A, knowledge graph relationships). Supabase powers auth, real-time collaboration, and document storage. ChromaDB stores the vector embeddings. The whole thing was built in 24 hours, and I learned a ton about embeddings, vector search, and streaming AI responses.
Thoughts
Building Nest showed me how much potential there is in AI-native tools that don't just augment existing workflows—they fundamentally rethink them. Dropbox is great for what it does, but the future of document management isn't just storage. It's understanding how your team's knowledge connects.