Language · 2025

VNASelf, a multi-agent assistant

A supervisor agent routes requests to specialist agents for calendar automation, OCR-based task capture and document questions, behind a real-time chat UI.

Supervisor + 3 specialist agents on LangGraph

Diagram of a supervisor agent routing to calendar, OCR and retrieval agents
Full write-up: problem, architecture, results, training, lessons. One-minute summary. Switch to Deep dive for the full write-up.
Problem
Personal admin is scattered across a calendar, photos of receipts and notes, and documents you cannot search. One chat window should be able to handle all three without a single monolithic prompt.
Approach
A LangGraph supervisor agent classifies each message and routes it to a specialist: Google Calendar automation, OCR-based task management (PaddleOCR / VietOCR for Vietnamese text), or a cloud-backed RAG agent over the user's documents on NeonDB + Supabase. FastAPI backend with WebSocket chat and per-thread history; React + Vite frontend.
Result
Working end-to-end assistant with real-time chat, chat-history threads and three specialist agents, containerised with Docker. [Add one measured number: e.g. routing accuracy on N test prompts, or OCR field accuracy on Vietnamese receipts.]

Problem

The usual single-prompt chatbot breaks as soon as a request needs a tool: creating an event, reading a receipt photo, or answering from a document it has never seen. The goal was a small system where each capability is its own agent with its own tools and prompt, and a supervisor decides who handles what, so adding a capability does not mean rewriting one giant prompt.

Architecture

Pipeline — select a block to read what it does

What I would show a reviewer

  • [Screenshot or short clip of the chat routing one message to the calendar agent and the next to the OCR agent.]
  • [An example receipt photo, the OCR output, and the task that was created from it.]
  • [A document question with the retrieved passage shown next to the answer.]

Challenges and lessons

  • Routing is the product. Most failures were the supervisor picking the wrong specialist, not the specialists failing. [Describe how you evaluated routing and what you changed in the supervisor prompt or state schema.]
  • Vietnamese OCR. PaddleOCR handles layout, VietOCR handles diacritics better on handwriting-like fonts; [note which one won on your receipts and why you kept both].
  • State across agents. Keeping one LangGraph thread per conversation meant the RAG agent could see what the calendar agent had just done. [Mention any limitation you hit.]

Code: github.com/x23d8/X23D8-muilti-agents (MIT). The README there is short; this page is the fuller write-up.