# Kepta: Legal Engineer task

This task is part of the application for both Legal Engineer roles at Kepta:

- **Legal Engineer, Workflows and Automation**
- **Legal Engineer, Knowledge and Quality**

Everything in it is fictional. The company, the people and the facts are invented for this exercise.

## The case

Fernbrook Labs sp. z o.o., a Warsaw software house, wants to engage a senior backend developer, Anna Kowal, on a B2B agreement. Read `brief.md` for what the client told us, and `baseline-agreement.md` for an agreement an AI assistant drafted from the brief, without legal review.

The question underneath both tasks: **will this relationship hold up as B2B under Polish law, or is it an employment relationship with a different name?**

## The rules

- Use any tools you like, including Claude, ChatGPT, Claude Code or Codex. We expect you to.
- We test what you add on top of the tools. We run your file against cases you have not seen, and we compare the result with what the tools produce on their own.
- In the interview we change one fact in the case and ask you to adapt your file live and explain why.
- Plan for about three hours. Send **one Markdown file (.md, at most 100 KB)** through the form on kepta.legal/careers.
- Write in English. Cite Polish law where it matters (article and act).
- Do not include personal data of real people.

## Workflows and Automation: build the workflow, then make the case for 10x

Your file contains three sections:

1. **Instruction file** (at most 400 words). Instructions we can paste into Claude or ChatGPT as the system prompt or project instructions. With them, the model must:
   - ask the intake questions that matter before drafting anything,
   - draft a B2B agreement with a low risk of reclassification as employment when the facts allow it,
   - refuse or escalate, and say why, when the facts point to employment, whatever the client asks for.
2. **10x memo** (at most one page). How a lawyer handles B2B agreements today, from intake to drafting, review, signature and keeping the relationship compliant over time. Which steps you would automate, which human checkpoints stay, which eval guards each step, and how you would show that quality for the customer did not drop.
3. **Iteration log** (about 150 words). Two things the model got wrong while you worked, what you changed, and a short before and after excerpt for each.

## Knowledge and Quality: build the eval that would catch it

Your file contains five sections:

1. **Eval cases** (at least ten). Each has an input (a clause or a fact), the expected judgment, a severity (critical, high, medium, low) and the legal basis. At least three must turn on how the work is actually performed, not on the wording of the agreement.
2. **Rubric.** Criteria an AI reviewer can apply to any B2B agreement together with its brief, ending in a verdict: low, medium or high reclassification risk, plus what must be escalated to a lawyer.
3. **Top five risks** in `baseline-agreement.md`, ranked, with the legal basis for each.
4. **What cannot be judged from the document alone**, and what the system should ask the client instead.
5. **Iteration log** (about 150 words). Two things the model got wrong while you worked, what you changed, and a short before and after excerpt for each.

## Suggested file layout

```markdown
# <Your name>: <role>

## 1. ...
## 2. ...
```

Good luck. We read every file.