Agentic Recruiter: Multi-Agent Resume Screener

An intelligent, multi-agent system that automates the first pass of resume screening using the Google Agent Development Kit (ADK) and Gemini.

---

🎯 Problem Statement

In high-volume hiring pipelines, the initial screening phase is often the biggest bottleneck. Recruiters face three core challenges:

1. Inconsistent Screening: Different recruiters (or the same recruiter on different days) may evaluate candidates differently, leading to bias and noise.
2. Hidden Gaps: Keyword matching misses the nuance. A candidate might match "Python" but lack the specific "real-time data processing" experience required for the role.
3. Slow Feedback: Hiring managers need detailed reasons why a candidate is a fit or not, but generating detailed reports for every applicant is impossible at scale.

💡 Solution

Agentic Recruiter is not just a chatbot; it's a structured multi-agent workflow designed to replicate the reasoning process of a human recruiter. Instead of a single "pass/fail", it produces:

• Structured profiles for both the Job Description (JD) and Candidates.
• A defensible, evidence-based fit score (0-100).
• A recruiter-ready report with interview focus areas.

It leverages Google's Agent Development Kit (ADK) to manage state, tools, and memory, ensuring a robust and observable pipeline.

🏗️ Architecture

The system is composed of five specialized agents orchestrated to perform a sequential screening workflow.

graph TD
    User((Hiring Manager)) -->|JD & Resumes| Root[Root Orchestrator - Agentic Recruiter]
    
    subgraph "Screening Pipeline (Sequential Agent)"
        Root -->|Parse JD| JD[JD Analyst Agent]
        JD -->|Job Profile| CP[Candidate Profiler Agent]
        CP -->|Candidate Profiles| FS[Fit Scorer Agent]
        FS -->|Scores & Evidence| RW[Report Writer Agent]
    end
    
    RW -->|Final Report| Root
    Root -->|Response| User
    
    subgraph "Tools & Memory"
        Tools[Document Parser - PDF/DOCX/URL]
        Memory[(Session & Long-term Memory)]
    end
    
    JD -.-> Tools
    CP -.-> Tools
    Root -.-> Memory

Loading

The Agent Team

Agent	Role	Model	Tools
JD Analyst	Parses job descriptions into structured requirements (skills, seniority, constraints).	Gemini 2.5 Flash-Lite	extract_job_profile, parse_document
Candidate Profiler	Extracts structured data from resumes (experience, skills, education).	Gemini 2.5 Flash-Lite	extract_candidate_profile, parse_document
Fit Scorer	Compares structured JD and Candidate profiles to compute a weighted score.	Gemini 2.5 Flash (Stronger reasoning)	compute_fit_score
Report Writer	Synthesizes scores into a ranked markdown table and summary.	Gemini 2.5 Flash-Lite	None (Synthesis only)
Root Orchestrator	Manages the user conversation, routes tasks, and handles memory.	Gemini 2.5 Flash	All sub-agents, load_memory

✨ Key Features

• Multi-Format Ingestion: Handles raw text, URLs, PDFs, and DOCX files automatically.
• Structured Reasoning: Uses JSON schemas to force the LLM to be rigorous about data extraction (vs. free-form hallucination).
• Long-Term Memory: Remembers hiring preferences and past roles across sessions using ADK's MemoryService.
• Observability: Integrated LoggingPlugin tracks every agent thought and tool call.
• Evaluation: Includes a built-in evaluation set (evalset.json) to measure relevance, accuracy, and helpfulness using adk eval.

🚀 Setup & Installation

Prerequisites

• Python 3.10+
• Google Cloud Project with Vertex AI API enabled.
• Google Cloud SDK installed and authenticated.

Installation

1. Clone the repository:

   git clone https://github.com/sobhanb-eth/agentic-recruiter.git
   cd agentic-recruiter

2. Install dependencies:

   pip install google-adk google-cloud-aiplatform PyPDF2 python-docx requests

3. Configure Environment: Create a .env file or set environment variables:

   export GOOGLE_CLOUD_PROJECT="your-project-id"
   export GOOGLE_CLOUD_LOCATION="us-central1"
   export GOOGLE_API_KEY="your-api-key" # Optional, if using API key auth

📖 Usage

Running the Notebook

The core logic is contained in Agentic-Recruiter.ipynb. You can run this in:

• Kaggle Notebooks: (Recommended) Use the "Google Cloud SDK" add-on for easy auth.
• Local Jupyter: Ensure you have gcloud auth application-default login set up.

Interactive Chat

The notebook provides a chat() helper function to interact with the agent.

# Start a session
await chat("Hello! I have a new role to screen for.", session_id="hiring_session_1")

# Provide JD
await chat("Here is the JD: https://company.com/careers/senior-ml-engineer", session_id="hiring_session_1")

# Provide Resumes
await chat("Here is the first resume: ...", session_id="hiring_session_1")

# Get Results
await chat("Please rank the candidates.", session_id="hiring_session_1")

Running Evaluations

To verify the agent's performance against the defined test cases:

adk eval resume_agent evalset.json --config_file_path=test_config.json

🛠️ Technical Details

Document Parsing

The parse_document tool acts as a smart router. It detects the input type (URL vs. Text) and file format (PDF vs. DOCX) to apply the correct extraction logic (PyPDF2 or python-docx) before passing clean text to the LLM.

Memory Architecture

• Short-term: DatabaseSessionService (SQLite) persists the immediate conversation state.
• Long-term: VertexAiMemoryBankService (in production) or InMemoryMemoryService (dev) stores high-level summaries (e.g., "User prefers candidates with Fintech background") to personalize future interactions.

☁️ Deployment

This agent is ready for deployment on Vertex AI Agent Engine.

1. Organize code into a resume_agent/ directory.
2. Define agent_engine_config.json.
3. Deploy with ADK CLI:

   adk deploy agent_engine \
       --project=$PROJECT_ID \
       --region=$LOCATION \
       resume_agent/

---

Built with ❤️ using Google ADK and Gemini.