Resume Screening Multi-Agent
Orchestrator plus extraction and screening agents turn resume emails into scored, rationale-backed rows in Google Sheets—no copy-paste shortlists.
Walkthrough
Video
End-to-end view of how the workflow behaves in practice.
GitHub
Enter your email to unlock the repo link — we'll log access for our records.
Architecture
How pieces connect
Email ingress through orchestration into specialized agents, ending in deterministic Google Sheets output your team can trust.
Step 01
Gmail
Inbound messages with resume attachments (PDF) trigger the workflow.
Step 02
Orchestrator
Validates payloads, dedupes, sequences jobs, and handles retries across agents.
Step 03
Extraction agent
Pulls structured candidate fields from the résumé PDF for downstream scoring.
Step 04
Screening agent
Scores the profile against your job criteria with an LLM and records concise rationale.
Step 05
Google Sheets
Deterministic row writes: scores, rationale, flags for review—no manual paste.
Foundation
Job rubric
Versioned criteria you control; the model follows it instead of improvising.
Audit trail
Logs, message IDs, and row keys so you can trace every score back to a source email.
Problem
What it helps with
At many companies, screening is still brute force: hundreds of résumés, one recruiter, and a lot of Ctrl+C / Ctrl+V into a spreadsheet. Criteria drift between people, strong candidates get buried, and time-to-hire suffers. The work is repetitive, but it has to stay consistent and auditable.
Behaviors
- Detects inbound emails that include a résumé attachment (typically PDF)
- Runs a multi-agent pipeline: an orchestrator coordinates dedicated extraction and screening agents
- Extracts structured candidate details from the résumé (skills, experience, education, and other fields you define)
- Scores each candidate against configurable job criteria using an LLM, with short rationale for transparency
- Writes scored results into Google Sheets through a deterministic mapping—fixed columns, validation, and review flags
- Surfaces edge cases for human review instead of silent auto-rejects
- We use this pattern in Kwanso’s own hiring today; the same approach extends to other workflows and stacks
Flow
How it works
- — 01Gmail is watched for new messages; attachments and metadata are handed to the orchestrator
- — 02The orchestrator dedupes by message/thread, validates file types, and sequences extraction → scoring
- — 03The extraction agent parses the PDF into a structured candidate profile your rubric can consume
- — 04The screening agent evaluates that profile against your job criteria and produces a score plus rationale
- — 05A sheet writer appends a row with stable column semantics (and optional links back to the email or Drive file)
- — 06Operational logging, retries, and human-in-the-loop review for borderline scores keep the system production-safe
Want this for your team?
We adapt triggers, approvals, and integrations to your environment — then ship for production discipline, not demos.
Next step
Something similar in your stack?
Tell us what systems and policies matter—we'll map a pragmatic path to a pilot.