Hire me Club – AI-Powered Candidate Screening & Ranking Agent

Executive Summary

Mist Avinya Technologies LLP designed and deployed a fully serverless Agentic AI recruitment platform on AWS, enabling Hire me Club to autonomously screen, score, and rank candidates against job descriptions — cutting time-to-shortlist from days to hours using Amazon Bedrock Agents and Claude 3.5 Sonnet.

About the Customer

Hire me Club is an SMB HR-technology company in the recruitment sector, delivering end-to-end talent acquisition — sourcing, screening, shortlisting, and onboarding — for mid-market employers across multiple verticals.

Business Context

With a lean internal team processing 800 to 1,200 CVs per open role, Hire me Club needed to scale screening capacity without proportionally increasing headcount. Manual CV screening consumed 60–70% of recruiters' working hours, creating a critical bottleneck.

Key Challenges

Recruiter burnout and turnover from repetitive, low-value screening
5–7 business day average time-to-shortlist — losing top candidates to competing offers
Inconsistent screening quality, introducing bias risk and compliance exposure
Inability to scale during peak seasons without temporary staff (+35% cost-per-hire)
Growing client SLA pressure for faster shortlist delivery

Project Goals & Objectives

Reduce recruiter workload, accelerate shortlist delivery, and standardize screening — while maintaining auditability and human oversight.

Accelerate Shortlisting

Reduce shortlist generation from 5–7 business days to under 24 hours for standard roles.

Improve Recruiter Productivity

Cut recruiter time spent on CV screening by at least 60%, redirecting effort to candidate engagement.

Standardize Screening Quality

Apply AI-based scoring and ranking to reduce inconsistency and improve compliance across recruiters.

Scale Hiring Operations

Support 3x peak hiring volume without additional headcount, and achieve 90%+ agent-to-recruiter alignment.

Solution Overview

A fully autonomous Agentic AI candidate-screening platform built on Amazon Bedrock Agents, integrated with the customer's Applicant Tracking System (ATS).

How the AI Agent Works

Retrieves the job description, required skills, and scoring rubric from DynamoDB
Fetches candidate CV keys from Amazon S3 for the role
Extracts structured text from PDF/DOCX CVs using Amazon Textract
Evaluates each candidate with Claude 3.5 Sonnet using chain-of-thought scoring
Writes ranked scores, justifications, and confidence back to DynamoDB
Triggers recruiter-dashboard notifications via Amazon SNS

Agent Tools (Lambda-backed)

get_job_description — active JD, skills, and scoring rubric
list_candidates — candidate CV keys for the posting
extract_cv_text — async Textract extraction (multi-page/scanned)
submit_candidate_score — persists score/rank/justification + SNS
ReAct reasoning loop with per-candidate tool calls and reflection
Human-in-the-loop review for low-confidence candidates

Foundation Model Selection

Claude 3.5 Sonnet on Amazon Bedrock was selected after evaluating four models across eight criteria: output quality, context window, reasoning depth, latency, cost, tool-use reliability, licensing, and AWS-native support.

Why Claude 3.5 Sonnet

Best-in-class tool-use reliability and multi-step reasoning
200K-token context window handles long CVs and detailed JDs in a single pass
Native Bedrock integration — no data leaves the customer AWS account

Alternatives Considered

Amazon Titan Text — lower reasoning quality for nuanced multi-criteria evaluation
GPT-4o — not AWS-native; data residency risk, no native Bedrock Agents integration
Mistral Large — inconsistent tool-call behaviour in multi-step agentic loops

AWS Services Used

A secure, scalable, serverless AWS architecture optimized for enterprise-grade agentic AI operations.

Amazon Bedrock Agents

Core orchestration layer — ReAct reasoning loop, tool invocation, session memory, human-in-the-loop interrupts, and end-to-end trace logging.

Claude 3.5 Sonnet

Agent reasoning model for multi-criteria candidate evaluation, chain-of-thought scoring, and justification generation.

Amazon Textract

Extracts structured text from PDF and DOCX CVs, including scanned documents, using asynchronous multi-page APIs.

Amazon DynamoDB

Stores job descriptions, scoring rubrics, candidate scores, agent decision logs, and audit trails with PITR.

AWS Lambda

Backs all four agent tools with dedicated least-privilege IAM roles, deployed in a private VPC subnet.

Amazon S3 · Textract · SNS · API Gateway

S3 stores CVs (versioning + Object Lock); SNS delivers shortlist notifications; API Gateway + Cognito secure the recruiter dashboard.

Security & Responsible AI

Security, governance, and responsible AI were integrated throughout the architecture.

Security Architecture

Private VPC deployment with VPC endpoints — no data on the public internet
Cross-account IAM roles with least-privilege, temporary STS credentials only
KMS customer-managed keys at rest; TLS 1.2+ in transit
CloudTrail across all regions; S3 Object Lock for log integrity
Amazon Cognito with MFA, SAML 2.0 federation for recruiter access

Responsible AI Controls

Bedrock Guardrails strip candidate PII before scoring to reduce demographic bias
Grounding checks route low-confidence candidates to human review
Every score includes a chain-of-thought justification recruiters must review
Bias monitoring with Amazon SageMaker Clarify (no disparate impact at 3-month review)
Per-posting session isolation prevents candidate data leakage between roles

Business Outcomes

Measurable improvements delivered within the first three months after deployment.

KPI	Baseline	Target	Actual Result
Time-to-Shortlist	5 to 7 Business Days	Under 24 Hours	4 Hours Average ↓ 94% Reduction
Recruiter Screening Time	65% of Working Hours	25% or Less	16% of Working Hours ↓ 75% Reduction
Agent-to-Recruiter Alignment	Manual Process	90% or Greater	93% Alignment ✓ Target Exceeded
Peak Hiring Capacity	1x Capacity	3x Without Headcount	3.2x Capacity ✓ Target Exceeded

Architecture — High Availability & Scalability

A fully serverless, AWS-native architecture focused on scalability, resilience, security, and operational simplicity.

AWS Architecture Diagram — Agentic AI Candidate Screening

High Availability

Bedrock Agents, Lambda, S3, DynamoDB, SNS, Textract, and API Gateway are inherently multi-AZ with no single point of failure in the agent execution path. RTO 1 hour, RPO 15 minutes via DynamoDB PITR and S3 versioning.

Scalability & Resilience

Lambda scales to 1,000 concurrent executions; Bedrock scales inference on demand; a Step Functions fan-out batches large candidate pools (groups of 20) — cutting processing time by 68% for high-volume roles.