AgentHiring AI Recruiting Server

AgentHiring is a state-of-the-art AI Recruiting Concierge Agent and multi-stage candidate discovery engine designed to streamline talent sourcing. It leverages the Google Agent Development Kit (ADK) to establish an interactive reasoning chat concierge, backed by the Model Context Protocol (MCP) server, and integrates a highly optimized offline candidate ranking pipeline with adversarial honeypot trap filtering.

Built for the Kaggle AI Agents: Intensive Vibe Coding Capstone using Google ADK and MCP.

---

🌟 Key Features

1. AI Recruiting Concierge (Google ADK): An interactive agent powered by gemini-2.5-flash that understands natural language commands (e.g. "Audit candidate CAND_0000002 for honeypot traps", "Compare top 3 matches for Python developer").
2. Model Context Protocol (MCP) Server: Exposes core recruiting tools (ranking, profile retrieval, honeypot audits, JD parsing) over stdio, allowing integration with clients like Cursor, Claude Desktop, or custom scripts.
3. Hybrid Sourcing & Ranking Pipeline: Combines lexical BM25 and dense semantic search (BAAI/bge-base-en-v1.5 on CPU) to scan and rank profiles.
4. Adversarial Honeypot Trap Filter: Detects and flags copy-pasted summary templates, chronological career alignment issues, and keyword-stuffed resumes (100% detection rate on candidate decoys).
5. Interactive Recruiter Dashboard: Premium dark-themed Streamlit application featuring candidate matching sliders, card expansion breakdowns, and a live AI Concierge Agent chat tab.

---

📸 Demo & Screenshots

Live AI Recruiting Chat Interface

Here is the interactive recruiting agent answering queries inside the Streamlit dashboard:

Sourcing & Interactive Playback

Here is a demonstration of the agent dynamically executing local MCP auditing and parsing tools:

---

⚙️ System Architecture

AgentHiring moves beyond static applicants tracking systems (ATS) by placing an intelligent reasoning loop on top of a highly optimized offline search engine:

Recruiter / Client
   │
   ├── (Natural Language Query) ──►  Google ADK Agent (talentlens_recruiting_concierge)
   │                                  │ (Decides which tools to run)
   │                                  ▼
   ├── (JSON-RPC stdio protocol) ──►  FastMCP Server (AgentHiring AI Recruiting Server)
   │                                  │
   │                                  ├── parse_job_description_tool
   │                                  ├── rank_candidates_tool
   │                                  ├── get_candidate_profile_tool
   │                                  └── detect_honeypot_trap_tool
   │                                  ▼
   └── (Optimized Engines) ────────►  BM25 Search + Vector Semantics + Honeypot Auditing

---

🚀 Setup & Installation

Prerequisites

• Python 3.10 or higher
• Google Gemini API Key (optional, for live AI chat interaction)

Steps

1. Clone & Install Dependencies:

   git clone https://github.com/mohd-ibadullah/AgentHiring.git
   cd AgentHiring
   pip install -r requirements.txt
   

2. Configure API Keys: Copy .env.example to .env and fill in your Gemini API Key if you want to use the live Gemini model:

   cp .env.example .env
   

3. Run Pipeline Setup (Model & Embeddings Cache): For first-time runs, pre-download the embedding models and compute candidate indices offline:

   • Windows: powershell -File setup.ps1
   • Linux/Mac: ./setup.sh

---

💻 How to Run

1. Launch the Streamlit Recruiter Dashboard

Launch the interactive web application which contains both the candidate discovery list and the Agentic chat panel:

streamlit run app/streamlit_app.py

2. Run the Interactive CLI Agent

Start a command-line chat session with the Recruiting Concierge Agent:

python run_agent.py

Or execute a single command directly:

python run_agent.py --prompt "check honeypot for CAND_0000002"

3. Start the MCP Server

To connect AgentHiring's tools to Cursor or Claude Desktop, start the protocol server:

python src/mcp_server.py

---

📊 Evaluation & Verification

To validate the ranking quality of AgentHiring, we include an automated evaluation module that compares our multi-stage pipeline against a standard BM25 Lexical Baseline.

You can run this evaluation script locally to verify the performance numbers:

python src/evaluate.py

Evaluation Metrics Summary (vs. BM25 Baseline)

The multi-stage pipeline yields substantial improvements over standard keyword-matching ATS:

Metric	Relative Lift (AgentHiring vs BM25 Baseline)	Rationale
Precision@10	+150.0% relative lift	Measures lexical-semantic alignment precision boost
Recall@20	+150.0% relative lift	Captures broader pool of relevant candidates
NDCG@10	+133.2% relative lift	Measures ranking sequence quality
Honeypot Rate (Top 1000)	100% Filtered (0.0% Ours vs 30.1% Baseline)	Stage 2 filters 301 decoy profiles from the BM25 pool

---

🧪 Running Tests

Verify the agent, MCP tools, and server integrations:

python -m unittest tests/test_agent.py

---

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.