Introspect AI

Overview

Introspect AI came from the idea that ChatGPT and Claude conversations are becoming a kind of personal cognitive archive. They capture how we think, what we are curious about, what we ask for help with, and how our interests evolve over time. I wanted to build a tool that made that archive visible.

The product transforms exported ChatGPT and Claude histories into an interactive 3D map. Each conversation becomes a point in space, embedded using semantic representations and compressed into 3D with UMAP. Similar conversations naturally cluster together, making it possible to visually explore patterns across thousands of AI interactions, almost like a Spotify Wrapped for your AI usage.

The system combines a React and Three.js frontend with a Flask backend that handles parsing, embedding, dimensionality reduction, clustering, labeling, and summarization. The pipeline supports both ChatGPT and Claude exports, generates embeddings through Voyage, reduces them into 3D using UMAP, clusters related conversations with KMeans, and uses Claude or Groq to create topic labels and summaries. Supabase and pgvector handle storage and semantic retrieval.

The interface was designed to feel exploratory rather than analytical. Users can zoom through their conversations, hover over points to inspect titles and content, search for concepts like “CUDA” or “economics,” filter by user, and summarize clusters. We also experimented with a social layer where users could “poke” someone’s brain and start a group chat around a specific cluster or conversation topic.

What made this project interesting was that it was not just another AI wrapper. It was a way to make AI usage itself observable. As people spend more time thinking through models, there is an opportunity to build tools that help them understand their own intellectual patterns, recurring interests, and shared curiosity with other people.

Technical Highlights

• ChatGPT and Claude export parsing
• Voyage embeddings for semantic representation
• UMAP dimensionality reduction into 3D space
• KMeans clustering with AI-generated topic labels
• React, Three.js, and WebGL interactive visualization
• Supabase and pgvector for storage and semantic search
• Groq and Claude-based summarization and topic exploration