AI trail: Darwinbox PM take-home
How AI was used to produce D1 (decision brief), D2 (build sheet) and D3 (prototype), what was checked, and what was rejected or hedged. ← Back to the prototype
In three lines. I worked in GitHub Copilot CLI, an agentic terminal assistant running Claude Opus 5.5. It read the case, ran web searches to check that the regulation I cited was current, drafted D1/D2 as HTML and rendered them to PDF, and built and browser-tested this prototype.
I used AI for speed, for breadth on the legal landscape and for code. The things a reviewer should hold me to are the framing, the recommendation, the numbers I assumed and the product boundaries.
Legal search results were treated as leads to verify, not as facts. That is why D1 §5 is written as “what I’d verify, and with whom”.
Tools
| Tool | Used for |
|---|---|
| GitHub Copilot CLI (Claude Opus 5.5) | Reading the case, structuring the argument, drafting D1/D2, writing the prototype code, running the checks below |
| Web search (inside Copilot CLI) | Checking the current status of regulation as of Oct 2026: EU AI Act timing, US state AI-in-hiring laws, India DPDP Rules |
| Headless Chrome | Rendering D1/D2 HTML to A4 PDF, then checking page count (≤2) and layout visually |
| Playwright (system Chrome) | An automated click-through test of every prototype route |
Workflow, step by step
| # | Step | What AI did · what I checked or decided |
|---|---|---|
| 1 | Brief | I gave the agent the recruiter’s email and the case PDF, and asked it to work out D1, D2 and D3 independently. It extracted the PDF text and listed the constraints: 2-page limits, a login-free D3, assumptions stated in one line each. |
| 2 | Context | It pulled in my own interview prep notes on Darwinbox’s published AI direction: Cortex, with governed agents and decision receipts. That is why the audit story is built around per-decision receipts. |
| 3 | Regulation currency check | It ran three web searches: the EU AI Act Digital Omnibus timing, US state laws (NYC LL 144, Illinois, Colorado, California) and India’s DPDP Rules phasing.
Hedged: the reported move of the Annex III high-risk deadline to Dec 2027 is written as “reported”, and the recommendation is to build to the high-risk standard regardless. Rejected: one result gave a wrong bill number for the Colorado AI Act, so no bill numbers or effective dates are cited for US states. They appear as items to verify with US counsel. |
| 4 | Framing | The agent and I worked through the core reframe: the customer is losing candidates to a slow “yes”, not a slow “no”. From there came the three options, the recruiter-hours model (assumptions stated in D1/D2) and the gates that govern autonomy. |
| 5 | D1 and D2 | The agent drafted print-styled HTML, rendered it to PDF, checked page count and inspected page images. It fixed a table split across a page break and used the spare space for a “what we tell the head of TA” paragraph. |
| 6 | D3 prototype | A single-file HTML app with no backend and synthetic data. An automated test ran all routes: Advance and booking; Review from an unclear knockout; India knockout auto-close followed by a human review request; Germany knockout routed to human confirmation; the recruiter decision requiring a reason code; a policy change blocked for Germany; a denied unlock request; criteria lint blocking proxies (age, postcode) and accepting a valid criterion; receipts export. Result: 0 JavaScript errors. |
Calls I am accountable for (not delegated to AI)
- Not shipping model-score auto-rejection, while still matching the competitor on speed. Autonomy is earned per country through gates.
- Germany: no automated adverse outcome until counsel and the works council agree. Starting the works council in week 1, because it is the critical path.
- Screening frontline roles with a structured mobile screener instead of CV-to-JD ranking. This is my hypothesis for why the hiring manager’s best hire would have been missed, and the backtest is designed to test it.
- The LLM extracts evidence, but a deterministic scorer assigns the band. That makes outcomes reproducible and explainable to an auditor.
- All numbers marked “assume” are mine and are labelled as such. They should be replaced by the customer’s data in week 3.
Limits of this AI-assisted work
- Legal points are hypotheses, not advice. The searches surfaced inconsistent details, which is exactly why each one has a named owner to verify it with.
- The prototype’s backtest and fairness figures are illustrative. They show the shape of the evidence an auditor would get, not a result.
- No customer or candidate data was used. The prototype collects nothing and sends nothing anywhere.