🕵️‍♂️ Item-Inspector AI

Visual AI for Product Condition Assessment & Human-like Reporting

Upload product images, let BLIP-2 understand the item, generate human-like condition reports with Phi-4, and enjoy the magic of zero-shot image-to-text reasoning. Also… there's a secret mini-game.

📚 Table of Contents

• Features
• Banner
• Project Structure
• Installation Guide
• Hardware & GPU Setup
• Bonus: Tic-Tac-Toe AI
• Technology Stack
• GitHub Topics
• License
• Contact

---

✨ Features

• AI Product Recognition – Detects object type: Watch, Shoe, Phone, etc.
• Material Identification – Metal, Leather, Glass, Suede? We got it.
• Visual Condition Tags – Custom per-item labels (like “scratched glass” or “torn strap”).
• Score Calculation – Evaluates product damage level and assigns a 4–10 score.
• Natural Language Report – Uses Phi-4 LLM to describe condition in ~50 human-like words.
• Frontend Upload UI – Drag, drop, analyze.

---

📸 Banner

---

🗂 Project Structure

Item-Inspector AI/
├── backend/
│   ├── app.py               # This FastAPI file
│   ├── requirements.txt
│   ├── python_gpu_test.py   # Check if TensorFlow, pytorch & numpy runs on GPU
├── frontend/
│   └── index.html           # Web UI for uploading images
├── sample_images/
│   └── example_watch.jpg    # Example test image
├── just_for_fun/
│   └── tic_tac_toe.py       # Tic-Tac-Toe AI game
├── README.md

---

🛠 Installation Guide

🔗 Prerequisites

• Python 3.10+ (recommended Python 3.10.11 for GPU usage on windows)
• GitHub Desktop or Git CLI
• Ollama installed & phi4(phi4:14b-q4_K_M) model downloaded

---

📥 1. Clone the Repo

git clone https://github.com/Rooshikesh/Item-Inspector-AI.git

cd Item-Inspector-AI/backend

---

📦 2. Create Virtual Environment

python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate

---

📦 3. Install Dependencies

pip install -r requirements.txt

---

🧠 4. Start Ollama with Phi-4

ollama run phi4:14b-q4_K_M

---

🚀 5. Launch FastAPI

uvicorn app:app --reload

Go to: http://127.0.0.1:8000/docs

---

🌐 6. Use Web Interface (Optional)

Open frontend/index.html in your browser. Drag and drop product images.

---

⚡ Hardware & GPU Setup

If you're planning to run BLIP-2 on GPU for maximum performance, keep the following in mind:

✅ Hardware Requirements

• NVIDIA GPU with at least 8–12GB VRAM
  • Recommended: RTX 3060 or higher
• CUDA-compatible drivers installed
  • Check GPU visibility with: nvidia-smi
• Python: Version 3.10+

✅ Python Environment for GPU

• Install PyTorch with CUDA support:

  pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
  

• Our code already includes:

  device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
  torch_dtype=torch.float16
  

  This ensures your models run on GPU if available.

✅ BLIP-2 Optimization Settings

• Make sure BLIP-2 loads with:

  device_map="auto", torch_dtype=torch.float16
  

• Images are correctly converted to RGB before inference:

  img = Image.open(file.file).convert("RGB")
  

🧪 Verify GPU with Our Utility Script

Run the included python_gpu_test.py file to confirm if TensorFlow, PyTorch, and NumPy are GPU-ready:

cd backend
python python_gpu_test.py

This script will print the detected GPUs, framework versions, and whether each is using the GPU or CPU.

---

🤖 Bonus: Tic-Tac-Toe AI

When you need a break from debugging and BLIP-2 hallucinations:

cd just_for_fun
python tic_tac_toe.py

• Supports easy, medium, and hard mode
• Uses Minimax algorithm in Hard mode to destroy your confidence 🔥

---

💡 Technology Stack

• BLIP-2 (Salesforce) - Vision Language
• Phi-4 (Ollama) - Language Generation
• FastAPI - Backend Framework
• HTML/JS - Minimal Frontend
• Hugging Face Transformers
• PyTorch

---

🏷️ GitHub Topics

ai, blip2, phi4, fastapi, transformers, computer-vision, image-classification, product-inspection, natural-language-generation, multimodal-ai, semantic-analysis, ecommerce-ai, repairtech, humanlike-ai, condition-scoring, pytorch, webapi, backend, frontend, python

---

📄 License

MIT — use it, share it, modify it. Just don’t forget to smile when it works.

---

✉️ Contact

Rooshikesh Bhatt rooshikeshbhatt@gmail.com

---