The Problem

Single-turn LLM chats can't handle complex, multi-step business logic. Most enterprise "AI integrations" end up as opaque wrappers: hard to debug, impossible to modify, and completely invisible to the people who need to trust them. There was no good tool for visually designing, testing, and deploying autonomous agent systems without writing infrastructure from scratch.

What I Built

Linea is a low-code/no-code environment where you architect AI workflows by connecting nodes. Each node is a discrete unit: an LLM call, a vector retrieval step, a tool invocation (Stripe, Slack, Google Search). You wire them together visually, backed by a reliable execution engine and a TypeScript SDK for embedding workflows in other products.

Technical Highlights

Node-Based Orchestration Built on React Flow and TypeScript. The canvas renders in real time as you build, propagating data contracts automatically so you catch mismatches before running anything. Agentic Memory Integrated a RAG pipeline using pgvector and FastAPI. Agents maintain long-term context across multi-step tasks by retrieving relevant memory chunks at each node execution. State Machine for Streaming Wrote a custom state machine to manage streaming LLM responses mid-workflow. Tool-use execution, where the model calls an API and waits for a result, is handled as a first-class state transition, keeping the UX smooth at sub-200ms response latency. Scalable Architecture Fully containerized with Docker. Stateless execution nodes scale horizontally; the state machine layer handles concurrency without workflow logic breakdown.

Stack

Next.js · NestJS · LangGraph.js · PostgreSQL · pgvector · Redis · BullMQ · Clerk

Result

• Reduced time to deploy custom business agents by ~70%
• Made agent logic auditable and editable by non-engineers
• Eliminated the "black box" problem: every decision in the pipeline is a visible, inspectable node