Legal help is one of the most expensive forms of expertise, and the gap between "I have a legal question" and "I can afford a lawyer" is where most people simply give up. Advocate explores how far modern AI can close that gap responsibly: a system that takes a plain-language description of a legal situation and produces structured analysis — relevant issues, applicable principles, and reasoned assessment — rather than a generic chatbot answer.

Architecturally, it pairs the Claude API for legal reasoning and natural-language interaction with neural network components for [classification of case type / outcome prediction — confirm the split]. Prompt engineering was a genuine engineering discipline here: legal analysis demands structure, citation discipline, and calibrated uncertainty, and getting a model to say "this is ambiguous and here's why" instead of confidently guessing took many iterations of prompt architecture and output validation.

The defining feature is the benchmark. I evaluated the system against real Supreme Court cases — feeding it case facts and comparing its reasoning and predicted outcomes against actual decisions. That created a measurable accuracy target instead of vibes, and it surfaced the hard truths: the system handles well-settled doctrine far better than novel questions, and confidence correlates poorly with correctness unless you explicitly engineer calibration in.

The struggles beyond the technical were ethical, and I treated them as design requirements: aggressive disclaimers, framing output as analysis rather than advice, and refusing the temptation to make the system sound more certain than it is. Advocate taught me that in high-stakes domains, the engineering challenge isn't capability — it's honesty about capability.