🏗️ Construction Manager Procurement Agent

An Intelligent Procurement Agent built using Google Agent Development Kit (ADK) and Gemini / LiteLLM, designed to manage construction material orders while enforcing site-specific governance rules, persistent memory, and human-in-the-loop approvals.

This project demonstrates how to build enterprise-grade, policy-aware agents with deterministic behavior, resumable execution, and clean separation between orchestration, tools, memory, and UI.

---

✨ Key Features

🔒 Persistent Site Memory

• Stores site-specific rules such as:

  • Approval limits
  • Vendor bans

• Backed by SQLite

• Survives agent restarts and conversations

🧠 Deterministic Procurement Logic

• Vendor selection via structured tools (no LLM guesswork)
• Strict enforcement of banned vendors
• Explicit material, quantity, and pricing checks

🛑 Human-in-the-Loop (HITL) Approval

• Orders exceeding approval limits automatically pause
• Approval / rejection handled declaratively via ADK’s require_confirmation
• Seamless pause → resume workflow

🔁 Resumable Agent Execution

• Uses ADK ResumabilityConfig
• Invocation resumes exactly where it paused after approval

🖥️ Interactive UI

• Streamlit-based chat interface
• Approval buttons for managers
• Live visualization of agent pauses and resumes

---

🧩 Architecture Overview

User (Streamlit UI)
        ↓
Supervisor Agent (ADK)
        ↓
────────────────────────────────
| Tools Layer                  |
| - Vendor Search              |
| - Approval / Confirmation    |
| - Memory Persistence         |
────────────────────────────────
        ↓
Persistent Memory (SQLite)

Agent Roles

• SupervisorAgent

  • Orchestrates procurement flow
  • Calls tools deterministically
  • Never leaks reasoning or system logic

• MemoryRecallAgent

  • Retrieves stored site rules
  • Uses ADK LoadMemoryTool

---

📁 Project Structure

omnagvekar-construction_manager_agent/
├── app.py                     # Streamlit UI + HITL handling
├── agents/                    # ADK agents and app definition
│   ├── Supervisor.py
│   ├── memory_agent.py
│   └── compiled_agent.py
├── tools/                     # Deterministic business logic
│   ├── vendor_search.py
│   ├── memory_tools.py
│   └── confirmation.py
├── core/                      # Model + settings
│   ├── llm.py
│   └── Setting.py
├── utils/                     # Database & session management
│   ├── db_manager.py
│   └── memory_db.py
├── prompts/                   # System & memory prompts
├── data/                      # Mock vendor data
│   └── mock_vendors.json
└── README.md

---

⚙️ Setup & Installation

1️⃣ Python Version

Python >= 3.13

Install uv

MacOS/ Linux:

curl -LsSf https://astral.sh/uv/install.sh | sh

Windows (PowerShell):

irm https://astral.sh/uv/install.ps1 | iex
# OR
pip install uv

2️⃣ Install Dependencies

pip install -e .

Create and Activate Virtual Environment and Activate the virtual environment

From the project root:

uv venv

# Windows
.venv\Scripts\activate

# Windows (Powershell)
.venv\Scripts\Activate.ps1

# Linux / MacOS
source .venv/bin/activate

Install all dependencies defined in pyproject.toml:

uv pip install -e .
# OR
uv sync

3️⃣ Environment Configuration

Create a .env file using .env.template:

LLM_MODEL=gemini/gemini-2.5-flash
LLM_API_KEY=your_api_key_here
DB_URL=sqlite+aiosqlite:///agent_memory.db
VENDOR_DATA_DIR=./data/mock_vendors.json

---

▶️ Running the Application

Run the Streamlit app from the project root:

streamlit run app.py

---

🧪 Example Workflow

1️⃣ Store Site Rules

User

For the Pune site, the approval limit is 40000 and avoid BadRock Cements.

Agent

Rules for Pune site have been updated.

---

2️⃣ Place an Order

User

Order 100 bags of cement for the Pune site.

• BadRock is excluded
• Cheapest valid vendor exceeds limit
• Agent pauses

System

Manager approval required.

---

3️⃣ Approve / Reject

• Click Approve Transaction → Order executes
• Click Reject Transaction → Order cancelled

---

🛠️ Tooling Highlights

Approval Enforcement

• Implemented via ADK FunctionTool(require_confirmation=...)
• No approval logic duplicated in prompts
• Tool layer owns governance

Vendor Lookup

• Purely deterministic JSON filtering
• Supports multiple materials per vendor

Memory Storage

• SQLite-backed rule storage
• Flexible schema (semantic key matching for limits)

---

🧠 Design Principles

• No chain-of-thought leakage
• Tool-first execution
• Model-agnostic design
• Clear separation of concerns
• Enterprise-safe defaults

---

📜 License

MIT License © 2026 Om Nagvekar

---

🎯 Purpose

This project was built as part of an AI Engineer / Agentic Systems assessment, showcasing how to design reliable, auditable, and policy-aware agents using the Google ADK ecosystem.