The whole company, in one search box.

Curacity Brain is an internal AI that searches across every system we use to work — Coda, Slack, Notion, HubSpot, and Fathom — and answers questions in plain language. Anyone at Curacity can ask, in their own words, and get a sourced answer in seconds.

What it does, in one paragraph

You ask a question. Brain finds the most relevant content across all five connected systems, sends those snippets to Claude, and streams back a written answer with citations. It runs in your browser, behind your Google login. Every employee gets the same access — no API keys, no setup, no waiting for someone else to do the search.

The pipeline

How retrieval works

Once a day (every four hours, actually), a sync job pulls fresh content from each connected system and breaks it into small chunks of about 1,200 characters. Each chunk gets converted into a 512-dimensional embedding — a mathematical fingerprint of its meaning — using Voyage AI's embedding model. Those embeddings are stored in a Supabase pgvector database.

When you ask a question, Brain converts the question into its own embedding and finds the chunks whose embeddings are mathematically closest. Top matches get sent to Claude as context. Claude reads them and synthesizes the answer.

Crucially, this means Brain finds chunks based on meaning, not keywords. Asking "what's our churn picture?" can surface a meeting transcript that says "GRR is tracking at 94%" even if neither phrase appears in the question.

Connected sources

What it costs

The whole stack runs on free tiers and pay-as-you-go embeddings. At the current sync cadence (~1M tokens/day) embeddings cost under a dollar a month. Claude usage is per-question — typical questions cost a few cents.

Privacy and access

Brain is gated behind Google sign-in restricted to curacity.com Google Workspace accounts. There's no public access. All data stays in our infrastructure: our Supabase, our Cloudflare account, our keys.

We do not send chat questions or answers to any third party beyond Anthropic (Claude) and Voyage (embeddings), both of which are commercial AI providers with standard data-handling agreements. No training is done on our data.

What's next

• Better citations. Each source card already links to the exact page or recording — we'll polish how those land in each system.
• Saved conversations. Right now history is per-session; persistent history is the obvious next step.
• Slack notifications. Optional weekly digest of "what changed in our knowledge base" — surfacing new docs, decisions, and meetings.
• More sources as we add tools. Easy to plug new ones into the existing sync architecture.