Recruitment — Screening & Interview Pipeline
Automate resume screening, skills matching, culture assessment, interview scheduling, and hiring manager notification. Replaces Lever/Greenhouse automation rules with a single declarative pipeline. ENFORCED IN CODE: - Contact details, SSNs and card numbers are redacted before any screening agent sees an application. - A separate reviewer checks each recommendation for reasoning based on a protected trait (age, gender, race, religion, disability, national origin, family status, location), and a gate stops the run if it finds any.
Open in the playground and play with it →
halo
// ============================================================
// Recruitment — Screening & Interview Pipeline
// PURPOSE: Automate resume screening, skills matching, culture
// assessment, interview scheduling, and hiring manager
// notification. Replaces Lever/Greenhouse automation
// rules with a single declarative pipeline.
//
// ENFORCED IN CODE:
// - Contact details, SSNs and card numbers are redacted before
// any screening agent sees an application.
// - A separate reviewer checks each recommendation for reasoning
// based on a protected trait (age, gender, race, religion,
// disability, national origin, family status, location), and a
// gate stops the run if it finds any.
// ============================================================
// --- Policies ---
policy CandidatePII : Redact(PII) & Block(PromptInjection)
policy CostGuard : MaxSpend(1.00)
// --- Tools ---
tool ATSAPI : REST("api.lever.co/v1/opportunities", apiKey: Env.LEVER_KEY)
tool CalendarAPI : REST("www.googleapis.com/calendar/v3", apiKey: Env.GOOGLE_CAL_KEY)
tool HRSlack : POST("hooks.slack.com/services/hiring-team")
// --- Store ---
tool RoleRequirements : VectorStore(Env.PINECONE_KEY, namespace: "job-descriptions")
tool PastHires : VectorStore(Env.PINECONE_KEY, namespace: "successful-hires")
// --- Struct ---
schema CandidateApplication {
candidate_id : String
name : String
email : String
role_id : String
role_title : String
resume_text : String
cover_letter : String
values : String
team_id : String
job_skills : Array<String>
min_experience : Number
salary_expectation : Number
}
schema CandidateDecision {
candidate_name : String
match_score : Number
skills_gap : Array
culture_fit : String
recommendation : String
interview_slot : String
}
// --- Agents ---
agent Screener [openai/gpt-4o-mini, temp: 0.0]
requires CandidatePII, CostGuard
agent CultureAssessor [anthropic/claude-sonnet-4, temp: 0.2]
equips PastHires
requires CandidatePII, CostGuard
agent BiasReviewer [openai/gpt-4o-mini, temp: 0.0]
system """
You audit hiring recommendations. Reply with JSON only:
{"protected_trait_used": true|false, "reason": "..."}.
protected_trait_used is true if any part of the reasoning relies on age,
gender, race, religion, disability, national origin, family status,
or where the candidate lives.
"""
requires CandidatePII, CostGuard
agent HiringCoordinator [openai/gpt-4o-mini, temp: 0.3]
equips RoleRequirements, CalendarAPI
requires CostGuard
// --- The Pipeline ---
@timeout(3m)
@retry(2, 5s)
pipeline ScreenCandidate(application: CandidateApplication) -> CandidateDecision:
application
// Step 1: Parallel assessment — skills match + culture fit
|* {
Screener.match_skills(
resume: context.data.application.resume_text,
role_id: context.data.application.role_id,
required_skills: context.data.application.job_skills,
years_required: context.data.application.min_experience
),
CultureAssessor.assess_fit(
cover_letter: context.data.application.cover_letter,
values_statement: context.data.application.values,
team_culture: context.data.application.team_id,
past_successful_hires: PastHires
)
}
// Step 2: Gate — stop here if skills match is too low
|? (context.data.match_skills.score >= 40)
// Step 3: Synthesize recommendation
|> HiringCoordinator.recommend(
skills: context.data.match_skills,
culture: context.data.assess_fit,
role: context.data.application.role_id,
salary_band: context.data.application.salary_expectation
)
// Step 4: Gate — a second agent checks the reasoning for bias
|> BiasReviewer.check(recommendation: context.data.recommend, culture: context.data.assess_fit) as bias
|? (context.data.bias.protected_trait_used == false)
// Step 5: Gate — stop here if the recommendation is "reject"
|? (context.data.recommend.decision != "reject")
// Step 6: Branch — auto-schedule top candidates
|? {
(context.data.match_skills.score > 85) => HiringCoordinator.schedule_interview(
tool: CalendarAPI,
candidate_email: context.data.application.email,
interviewer_id: context.data.recommend.interviewer,
duration: "45m",
type: "technical"
),
else => System.Value(context.data.recommend),
}
// Step 7: Notify the hiring manager
|> HRSlack.post(
message: "New candidate screened",
candidate: context.data.application.name,
role: context.data.application.role_title,
match_score: context.data.match_skills.score,
culture_fit: context.data.assess_fit.rating,
recommendation: context.data.recommend.decision
)
// Step 8: Human decision
|! yield(
HRSlack,
context.data.recommend,
[context.data.match_skills, context.data.assess_fit],
48h
)
~> listen(HRSlack.Reaction)
// Step 9: If approved, push to ATS
|> ATSAPI.post(
event: "CANDIDATE_ADVANCED",
candidate_id: context.data.application.candidate_id,
stage: "interview",
recommendation: context.data.recommend
)