HR: Applicant screening with a human decision

examples/60-hr-screening.halo policies |* parallel |! people

Scores each application against the role, with contact details redacted before the model sees them, then asks a recruiter to approve the shortlist before it's shared with the hiring team. The model only recommends; a person decides who moves forward. Note: automated candidate scoring is regulated in some places (e.g. NYC Local Law 144, the EU AI Act). Keep the human approval.

Open in the playground →

// ============================================================
// HR: Applicant screening with a human decision
// ============================================================
// Scores each application against the role, with contact details
// redacted before the model sees them, then asks a recruiter to
// approve the shortlist before it's shared with the hiring team.
// The model only recommends; a person decides who moves forward.
//
// Note: automated candidate scoring is regulated in some places
// (e.g. NYC Local Law 144, the EU AI Act). Keep the human approval.
// ============================================================

policy CandidatePrivacy : Redact(Email) & Redact(Phone) & Redact(SSN) & Block(PromptInjection)
policy Fair             : Require(JSON)
policy Budget           : MaxSpend(0.50)

tool HiringChannel : Slack(Env.SLACK_BOT_TOKEN)

agent Screener [anthropic/claude-haiku-4-5, temp: 0.0]
  system """
    You screen job applications against a role description. Judge only
    job-relevant skills and experience. Ignore names, age, gender,
    nationality, photos, and anything else not about the work.
    Answer with JSON only.
  """
  requires CandidatePrivacy, Fair, Budget

agent Recruiter [anthropic/claude-sonnet-4-5, temp: 0.2]
  system "You turn scored applications into a short, fair shortlist with reasons."
  requires CandidatePrivacy, Budget

@timeout(10m)
@retry(2, 5s)
pipeline ScreenApplicants(role: String, applications: List, shortlist_size: Number) -> Result:
  applications
    // 1. Score every application against the role, two at a time
    |* each(application in applications, concurrency: 2) => Screener.score("""
         Role: {{role}}
         Score this application from 0 to 100 for fit. Reply as JSON:
         {"applicant_id": "...", "score": 0, "strengths": ["..."], "gaps": ["..."]}
       """, application) as scores

    // 2. Draft the shortlist
    |> Recruiter.shortlist(
         "Pick the top {{shortlist_size}} applicants for {{role}} and give one line of reasoning each.",
         context.data.scores
       ) as shortlist

    // 3. A recruiter approves before anyone is contacted
    |! yield(msg: "Share this shortlist with the hiring team?", [context.data.shortlist, context.data.scores]) as decision
    |? (context.data.decision.approved == true)

    // 4. Share it
    |> HiringChannel.post(channel: "#hiring", text: "Shortlist for {{role}}:\n{{shortlist}}")
    |> System.Print(context.data.shortlist)