Project
Automata
An autonomous AI agent that builds, verifies, and updates n8n workflows through natural language.
- Next.js
- PostgreSQL
- Gemini
- Laminar
Overview
Automata helps developers and automation engineers turn an idea into an n8n workflow without manually configuring every node and connection. Users describe what they want to automate, and the application retrieves relevant documentation, generates workflow JSON, checks its structure and logic, and deploys it to a connected n8n instance.
It supports both creating workflows and modifying existing ones through a conversational interface.
My contribution
I designed and built the entire application, from the chat interface and authentication to the AI pipeline, database, and n8n integration.
The pipeline classifies requests, extracts structured intent, retrieves node documentation using PGVector, and generates workflows with Gemini. I implemented deterministic structural checks, an AI critic for logical review, and bounded correction loops that feed validation errors back into the generator.
I also built workflow creation and update handling through n8n’s API, conversation persistence, and streamed progress updates so users can follow each stage.
Technical notes
Giving generation defined checkpoints:
I separated workflow generation into stages instead of relying on a single prompt. A deterministic validator checks for duplicate node names, dangling connections, and disconnected nodes before an AI critic reviews whether the workflow fulfills the requested intent.
When validation fails, the generator receives the existing JSON, the specific error, and the original context to attempt a correction. This makes failures easier to trace and gives retries a concrete problem to address, though additional model calls increase latency.
Grounding generation in documentation:
I implemented per-node retrieval with PGVector so the generator receives documentation for each predicted integration. This reduces the risk of a single similarity search overlooking a necessary node, at the cost of additional database queries.
The key lesson was that successful deployment does not prove correct execution. My next priority is evaluating retrieval quality and testing generated workflows against controlled service responses to establish a measurable reliability baseline.
Published on
Back to projects