MIRROR - AI-Powered Portfolio

Portfolio & knowledge base for Michail Berjaoui, Lead ML/LLM Engineer. Built with Flask, Qdrant, BGE-M3, and Phi-4.

Features

• Portfolio Landing - Dark cyber-violet showcase of CV, skills, experience
• AI Chat (RAG) - Ask questions about uploaded documents with source citations
• Web Scraper - Input a URL, scrape it, ask questions about the content
• Document Upload - Drag & drop PDF, DOCX, TXT, MD files for indexing
• Articles - Markdown-based technical blog
• Model Manager - Load/unload/switch LLM models at runtime
• Technical Choices - Full documentation of architecture decisions

Tech Stack

Component	Technology
Backend	Flask 3.1, Python 3.11+
LLM	Phi-4 14B Q4_K_M via llama-cpp-python
Embedding	BGE-M3 via sentence-transformers
Vector DB	Qdrant (Docker, HNSW + INT8 quantization)
PDF Parser	PyMuPDF
Scraper	trafilatura + BeautifulSoup

Quick Start

1. Start Qdrant

docker run -d -p 6333:6333 -p 6334:6334 \
  -v qdrant_data:/qdrant/storage \
  qdrant/qdrant:v1.12.4

2. Download a GGUF model

Place a .gguf model in the ./models/ directory:

mkdir -p models
# Example: download Phi-4 Q4_K_M
# huggingface-cli download microsoft/phi-4-gguf phi-4-Q4_K_M.gguf --local-dir ./models/

3. Install dependencies

pip install -r requirements.txt

4. Run

python run.py

Open http://localhost:5000

5. Load models

Click the ⚙ button in the top-right corner to:

1. Connect to Qdrant
2. Load the embedding model (BGE-M3)
3. Load the LLM (select a .gguf file)

Docker Compose

docker compose up -d

This starts both Qdrant and the Flask app.

Infrastructure

Optimized for 64 GB RAM, 12 CPU cores, no GPU:

• LLM: 10 threads, ~9 GB RAM
• Embedding: 2 threads, ~2 GB RAM
• Qdrant: 2 CPUs, ~4 GB RAM
• Remaining: OS + headroom

See docs/choix_techniques.md for full technical documentation.

Project Structure

MIRROR/
├── app/
│   ├── __init__.py          # Flask app factory
│   ├── config.py            # All configuration dataclasses
│   ├── routes/              # Flask blueprints
│   │   ├── main.py          # Landing page
│   │   ├── chat.py          # RAG chat API
│   │   ├── documents.py     # Document upload/management
│   │   ├── scraper.py       # Web scraping API
│   │   ├── articles.py      # Articles API
│   │   └── models_route.py  # Model management API
│   ├── services/            # Business logic
│   │   ├── embedding.py     # BGE-M3 embedding service
│   │   ├── llm.py           # Phi-4 LLM service
│   │   ├── qdrant_store.py  # Qdrant vector store
│   │   ├── rag.py           # RAG pipeline with citations
│   │   ├── scraper.py       # Web scraping
│   │   └── pdf_parser.py    # Document parsing + chunking
│   ├── static/css/style.css # Dark cyber-violet theme
│   ├── static/js/app.js     # Global JS (model manager, toasts)
│   └── templates/           # Jinja2 templates
├── articles/                # Markdown articles
├── docs/choix_techniques.md # Technical choices document
├── models/                  # GGUF model files
├── uploads/                 # Uploaded documents
├── docker-compose.yml
├── Dockerfile
├── requirements.txt
└── run.py

License

Private portfolio project by Michail Berjaoui.