Overview

What would a real AI-native second brain look like if it were built for understanding, not just storage?

A lot of knowledge tools are good at helping people store information, but much worse at helping them actually use it later. People save PDFs, screenshots, notes, research, and documents across different places, but once that information piles up, it becomes hard to revisit, connect, and learn from. I built Continote around that gap — an AI-powered knowledge workspace designed to ingest personal knowledge across multiple formats, structure it through a retrieval pipeline, and let users interact with it through semantic search, grounded Q&A, summaries, quizzes, and more agentic workflows.

Problem

People already have knowledge. The problem is that it becomes static the moment it gets stored.

Important information ends up spread across PDFs, notes, screenshots, images, research documents, and personal writing. Once that happens, it becomes harder to find the exact part you need later, connections across files get lost, reviewing material becomes manual and repetitive, dense content stays dense, and stored information becomes passive instead of useful. I wanted to build something that went beyond 'file upload + chatbot' and instead made personal knowledge feel alive again.

Approach

How do I turn scattered personal content into a system that can be searched, understood, and actively worked with?

Personal knowledge is messy — it doesn't live in one clean format, and it's not naturally structured for retrieval. I designed Continote as a layered workflow: ingest uploaded content across formats, extract usable text from documents and image-heavy sources, chunk content into retrievable segments, generate embeddings and store them in a vector index, retrieve relevant context from the user's own data, use grounded generation to answer questions, and extend that into agentic workflows that can synthesize, organize, and help users learn from their material.

Pipeline

Preprocessing, chunking, and retrieval quality were just as important as the generation layer itself.

Once a user uploads content, the system processes it into chunks that preserve enough context to stay meaningful while still being precise enough for retrieval. Those chunks are embedded and indexed so the workspace can support semantic search instead of relying on exact keyword matching. From there, I used a RAG-style architecture so answers stayed grounded in the user's own content rather than generic model behavior. That grounding mattered a lot — trust is everything in a knowledge product.

Agentic

The system doesn't just retrieve information when prompted. It helps users process, transform, and reinforce it.

I didn't want the system to stop at 'ask your documents a question.' I wanted it to begin acting more like an intelligent knowledge workspace that could do work with the user's information — generating summaries from dense material, creating quizzes from uploaded notes, synthesizing insights across multiple sources, and helping structure scattered knowledge into more usable outputs. That agentic direction pushed Continote toward a future where the workspace can participate in knowledge work, not just respond to it.

Design

The goal was to make advanced AI infrastructure feel like a natural extension of thinking.

A lot of AI knowledge tools feel technically impressive but operationally noisy. I wanted Continote to feel much simpler than the backend actually is. The experience was designed around a straightforward mental model: upload your knowledge, ask what you want to know, get grounded answers, generate useful outputs from your material, return later and continue building on it. The backend involves extraction logic, chunking strategies, embeddings, vector search, and agentic orchestration — but the product only becomes valuable if the user experiences it as clarity rather than complexity.

Reflection

Can software help people build an evolving relationship with what they know?

Continote reflects a bigger shift in how I think about software. A lot of applications are built around structured data that already exists in usable form. This project pushed me to think about how much personal knowledge is actually fragmented, hidden in mixed formats, and difficult to operationalize. What interested me most was not just storing that information, but building a system that could help users continuously interact with it — treating AI less like a flashy feature and more like an infrastructure layer for understanding.