iknowbrains

About the project

Upload a company's documents and its brand details. Ask questions about them in chat, and have it write and post social media content in the brand's voice.

iknowbrains is a side project with Jay Prakash Pathak and Krishna Mahato: an assistant that knows a company. It answers questions from the company's own documents, keeps its brand information in one place, and uses that to make and post social media content. A website widget and WhatsApp were planned too.

My parts were data_brain, a document question-answering server I built alone; the AI layer of odin, our server; and two dashboard UIs.

How it works

1. 1

   Upload

   Add text or PDF files.

2. 2

   Split and embed

   Files are cut into chunks and turned into vectors.

3. 3

   Ask

   Questions are answered from the closest chunks, with the last few messages as context.

4. 4

   Stream

   Answers, and tool calls, stream back as they're written.

5. 5

   Post for the brand

   The brand's details guide social posts, written and published from the same place.

Under the hood

data_brain

Express, LangChain and OpenAI embeddings, with the index rebuilt at startup.

odin

Koa and OpenAI, with streaming tool calls.

Dashboard

Next.js 15, tRPC, shadcn/ui and Recharts.

What I did

• data_brain, solo: upload text or PDF files, split and embed them, and answer questions with chat history
• The AI layer for odin, on Jay's server: normal and streaming answers, including streamed tool calls
• Two dashboard UIs on Jay's base: analytics, documents, projects, reports, search, team, AI agents, API keys and usage

Problems I hit, and how I fixed them

1. 01

   The in-memory search index was lost on restart.

   Fix Save document details to disk and rebuild at startup.

2. 02

   Only plain text worked.

   Fix A parser for each file type, starting with PDF.

Screenshots