ProcureAI – Intelligent RFP Management System

ProcureAI is a full-stack, AI-powered procurement dashboard.
It streamlines the RFP lifecycle by using LLMs to:

• Automatically generate data schemas from natural language requests
• Extract structured data from unstructured vendor PDF proposals

---

🛠️ Project Setup

Prerequisites

• Node.js (v18 or higher)
• Python (v3.10 or higher)
• PostgreSQL Database URL (e.g., Neon.tech or Supabase)
• Google Gemini API Key (Free tier via Google AI Studio)
• Ethereal Email Account (For testing SMTP)

---

1. Installation Steps

Backend

cd backend
python -m venv venv
# Activate Venv:
# Windows: venv\Scripts\activate
# Mac/Linux: source venv/bin/activate
pip install -r requirements.txt

---

Frontend

cd frontend
npm install

---

2. Configuration (.env)

Create a file named .env in the backend/ folder with the following variables:

# Database Connection (PostgreSQL)
DATABASE_URL="postgresql://user:password@host/dbname?sslmode=require"

# AI Provider (Google Gemini)
GEMINI_API_KEY="AIzaSy..."

# Email Settings (Ethereal.email)
MAIL_USERNAME="your_ethereal_email@ethereal.email"
MAIL_PASSWORD="your_ethereal_password"
MAIL_FROM="system@procureai.com"
MAIL_PORT=587
MAIL_SERVER="smtp.ethereal.email"

---

3. How to Run Locally

Step 1: Start Backend

cd backend
python create_tables.py  # Run once to initialize DB
uvicorn main:app --reload

Server runs at: http://127.0.0.1:8000

Step 2: Start Frontend

cd frontend
npm run dev

UI runs at: http://localhost:5173

---

4. Seed Data (Optional)

To quickly populate the database with test vendors for the demo:

cd backend
python seed_data.py

---

💻 Tech Stack

Component	Technology	Reasoning
Frontend	React + Vite	Fast build times and component-based architecture
UI Library	Material UI (MUI)	Enterprise-grade components (DataGrid, Cards) out of the box
Backend	FastAPI (Python)	High performance, async support, and native ecosystem for AI libraries
Database	PostgreSQL	Reliability of SQL for Users/Vendors + Flexibility of JSONB for dynamic RFP schemas
AI Provider	Google Gemini 1.5 Flash	Fast inference, large context window for PDFs, and cost-effective (Free tier)
PDF Parsing	pypdf	Lightweight library to extract raw text layers from PDF documents
Email	FastAPI-Mail + Ethereal	Robust SMTP implementation with a safe testing sandbox

---

📚 API Documentation

1. Create RFP (AI-Powered)

• Endpoint: POST /rfps/

• Description: Analyzes natural language prompt to generate a JSON Schema.

• Body:

  {
    "title": "Laptops",
    "prompt_text": "I need 10 Macbooks..."
  }

• Response: Returns RFP object with AI-generated json_schema.

---

2. Upload Proposal (AI-Extraction)

• Endpoint: POST /rfps/{id}/proposals/
• Description: Ingests a PDF, extracts text, and uses AI to map data to the RFP schema.
• Body: multipart/form-data (File)
• Response: Returns extracted data JSON (Price, Specs, Warranty) + AI Fit Score.

---

3. Manage Vendors

• Endpoint: GET /vendors/ | POST /vendors/
• Description: CRUD operations for vendor contact management.

---

4. Send Invites

• Endpoint: POST /rfps/{id}/send/
• Description: Triggers SMTP email dispatch to selected vendors.
• Body:

  {
    "vendor_emails": ["vendor@test.com"]
  }

---

🧠 Decisions & Assumptions

Key Design Decisions

• Hybrid Database Schema:
  Instead of creating a new SQL table for every RFP type (which is impractical), a JSONB column (extracted_data) is used. This allows storing diverse fields like Screen Size for monitors or Catering Menu for events in the same table without migrations.
• Schema-Guided Extraction:
  The LLM is guided by the specific JSON Schema generated during creation, ensuring strictly structured outputs that align with frontend table columns.
• Scoring Logic:
  The AI acts as a "Judge," comparing semantic meaning of the User Request vs. Vendor Proposal to assign a 0–100 Fit Score.

Assumptions

• PDF Format: Assumes PDFs contain selectable text layers (not scanned images/OCR required).
• Email Workflow: For this MVP, proposals are uploaded manually. In production, this would be replaced by an Inbound Email Webhook (e.g., SendGrid Inbound Parse).
• Single User: Currently supports a single Procurement Manager workflow (authentication out of scope).

---

🤖 AI Tools Usage

Tools Used: Large Language Models (Gemini/ChatGPT) & Cursor/VS Code

1. Boilerplate & Speed

   • AI scaffolded initial FastAPI CRUD endpoints and React components, saving hours of manual typing.

2. Debugging Complex Logic

   • Resolved CORS errors between React (Port 5173) and FastAPI (Port 8000).
   • Fixed React Hook dependency warnings in useEffect by refactoring data loading logic.

3. Advanced Prompt Engineering

   • Challenge: Getting AI to reliably output clean JSON without markdown formatting.
   • Solution: Iterated on system prompts with strict instructions:
     "You are a data extraction engine... Normalize all numbers... Return JSON ONLY."

4. Learning Outcome

   • Learned how to apply Zero-Shot NER (Named Entity Recognition) to extract specific entities (Price, Warranty) from unstructured text without training a custom model.