Design a System That Generates Personalized Recruiter InMails at Scale

Recruiters use LinkedIn Recruiter to find candidates who match a job's requirements among more than one billion LinkedIn members, and a single recruiter may need to send InMail messages to hundreds of candidates a day. Writing a personalized message for each candidate by hand takes a lot of time. A generic template is efficient but lacks personalization, and weak personalization can reduce candidates' willingness to reply and hurt overall recruiting results. Design a system that helps recruiters efficiently generate personalized InMail messages based on each candidate's information and the job's requirements. ```hint Which details are worth mentioning A profile holds many facts, but only a few make a message feel written for that person and relevant to this job. Decide how the system picks them. ``` ```hint Wrong is worse than generic Think about what happens when a draft states something about the candidate that is not true, and where in your design you would catch it. ``` ```hint Hundreds of drafts a day Look at which inputs change from one candidate to the next and which stay the same across a recruiter's outreach for one job, and what that means for cost. ``` ### Constraints and Clarifications - The candidate pool has more than one billion members, but a message is needed only for candidates a recruiter decides to contact. - A single recruiter may send InMail to hundreds of candidates a day. - The inputs are the candidate's information and the job's requirements; the recruiter is the sender of each message. ### Clarifying Questions - Does the system start from candidates the recruiter has already selected, or is finding candidates in scope too? - Must the recruiter review every draft before it is sent, or may drafts be sent in bulk? - Which profile data may the system use: only what the recruiter can see, and under which privacy settings? - What defines success: reply rate, positive replies, recruiter time saved, or hires? - Should messages follow the recruiter's own tone, saved templates or company branding? - Must a draft appear interactively for one candidate, in a batch for a whole list, or both? - Which languages must be supported? ### What a Strong Answer Covers - Clear framing: inputs, outputs, the recruiter's role, and a success metric tied to candidate replies rather than message volume - A step that selects relevant, verifiable personalization points from the profile and the job before any text is generated - A generation approach with its quality and cost trade-offs, such as template filling versus a generative model, or a large versus a smaller fine-tuned model - Factual grounding and safety checks: no invented claims, no sensitive or protected attributes, and spam controls - Training and feedback data from recruiter edits and candidate replies, including their biases - Offline evaluation and an online experiment design with guardrail metrics - Serving for hundreds of drafts per recruiter per day: latency, cost, caching and fallbacks ### Follow-up Questions - Reply rates depend heavily on how attractive the job is and how open the candidate is to moving. How would you measure the effect of the personalization itself? - How would you stop the model from using or implying protected attributes such as age, gender or ethnicity? - If recruiters start sending drafts without reading them, what changes in your safeguards? - How would you adapt drafts to each recruiter's writing style without leaking one recruiter's content into another's messages?