How I Built a Legal Assistant (Without Being a Lawyer) and Lived to Tell the Tale

This talk is a behind-the-scenes look at how I built a legal assistant platform that helps lawyers draft legal documents, track case progress, and get AI-powered guidance through contextual chat — without being a lawyer myself.

The real challenge was not technical at first — it was understanding legal workflows. I started by becoming a legal assistant to my wife (a practicing lawyer), iterating on real use cases, and translating them into code.

From a technical standpoint, the platform is powered by:

• LangGraph, which we chose over other agentic frameworks due to its high-level control of input/output flow, critical in legal contexts where traceability and decision logic matter.
• Gemini AI, selected for its large context window, after initially prototyping with ChatGPT agents. Legal use cases demand richer context, and Gemini proved to be more suitable for multi-agent orchestration.

The architecture includes Dockerized microservices, a backend-for-frontend layer, and a case service that communicates via RabbitMQ with a graph-processing engine. This engine runs REST-based LangGraph workflows that drive our LLM-powered assistants.

In the demo, I’ll show:

• How a case is processed end-to-end using agents, document ingestion, and LangGraph workflows
• The evolution of the stack and what technical decisions were driven by challenges like latency, model output storage, and managing LLM usage at scale
• Why modularity and context-awareness were vital in legal AI — and how we solved it
• What we’re building next: case-specific assistants, chat with real-time document context, RAG-enhanced flows, and even self-hosted model support for private legal deployments

We’re even training a junior dev from scratch on React and building a new interface together — tinkering at every level.