You already know: embeddings and two-tower retrieval (D4), learning to rank (D7), fairness testing (D10 and the credit case) and, from the e-commerce case, that hard constraints belong outside learned models. This case adds a two-sided market with capacity: a match needs the seeker to apply and the recruiter to respond, popular jobs saturate while others starve, descriptions are free text in a thousand dialects, and hiring is regulated. The design questions: what success means for both sides, how to understand jobs and people, how to spread attention, and how to stay fair and safe.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud. Employment law differs by country; this lesson describes common patterns, not legal advice.
1. The prompt, and the questions that change the design
A mid-level answer: "embed profiles and jobs and recommend the jobs people are most likely to apply to". A staff candidate first asks whose success counts (an application, a recruiter reply, a hire), how jobs and profiles are described, how attention is spread across jobs, and what legal and trust limits apply. Those answers produce a mutual objective, a skills taxonomy with embeddings, congestion control, and fairness audits with scam filtering.
Assumptions you state: