Build an AI legal review agent from scratch with zero instructions.
In-house legal teams are drowning in repetitive reviews for low-stakes contracts like NDAs and standard order forms. An AI agent that flags deviations from standard boilerplate saves hours of highly paid legal time. Building this project proves you can drop into a domain you know nothing about, find your own test data, design a safe human-in-the-loop workflow, and ship an evaluated product without asking for permission.
The Brief
You are presented with a real-world scenario from a high-paying job application: a small legal team manually reviews a high volume of low-stakes commercial contracts that are 80% identical. The prompt is simply: "Show us how AI could help with this." You are given no sample contracts, no required output format, and no specific feature requests.
Your task is to take total ownership. You must source realistic legal contracts yourself, define what "review" actually means in this context, and decide where the AI makes decisions versus where it surfaces information for a human lawyer. You will build a prototype that accepts a contract and outputs a usable review, complete with an evaluation framework proving your output is accurate and trustworthy.
The Idea Behind It
PortSwigger wants roving, hands-on builders to embed in non-technical departments for two-week sprints to ship working AI solutions. For their application task, they intentionally provided a vague legal scenario to test if candidates would wait for instructions or immediately source their own data, define their own scope, and build something useful.
This project prepares you for high-impact AI roles by forcing you to exercise extreme agency. It proves you don't need a product manager to feed you requirements; you can identify business friction, architect a safe AI workflow, and measure its actual impact independently.
What You Will Build
- A sourced dataset of realistic, low-stakes commercial contracts to test your system against.
- A written scope document defending your architectural decisions and defining strict boundaries for the AI.
- A working prototype that ingests a contract and outputs a structured review (e.g., flagging non-standard clauses).
- A human-in-the-loop workflow design ensuring a lawyer always makes the final call.
- A quantitative evaluation pipeline that measures the accuracy and time-saving impact of the agent's output.
Sourcing domain-specific data and designing a rigorous evaluation for high-risk text (legal) requires mature engineering judgment.
Build high-agency products that command £100k+ salaries in the AI Native Builder cohort.
Apply to Cohort