🧬 Enterprise AI Talent Platform

High-Dimensional Semantic Resume Analysis & RAG Orchestration Engine

The Enterprise AI Talent Platform is not just a resume parser—it is a sophisticated RAG (Retrieval-Augmented Generation) system. By projecting candidate profiles and job descriptions into a shared high-dimensional latent space, it performs semantic alignment far superior to traditional keyword matching. It features an interactive AI Interrogator that allows recruiters to "chat" with resumes to verify claims using evidence-based vector retrieval.

---

🏗️ System Architecture

The solution implements a Multi-Stage Reasoning Pipeline fusing deterministic logic with probabilistic Generative AI.

graph TD
    subgraph "Ingestion Layer"
    A[PDF/DOCX Upload] -->|PyMuPDF/Python-Docx| B(Text Extraction)
    B -->|Spacy NLP| C(Tokenization & Cleaning)
    end

    subgraph "Vector Compute Layer"
    C -->|all-MiniLM-L6-v2| D[Dense Vector Embeddings]
    D -->|IndexFlatL2| E[(FAISS Vector Store)]
    end

    subgraph "Inference Engine"
    F[Job Description] -->|Embedding| G(Semantic Similarity Calc)
    D <--> G
    B -->|Zero-Shot Prompting| H[Gemini 1.5 Flash]
    H -->|JSON Parsing| I(Structured Data: Exp, Edu)
    end

    subgraph "RAG Interrogator"
    J[User Question] -->|Vector Search| E
    E -->|Top-K Chunks| K[Context Window]
    K --> H
    H -->|Hallucination-Free Answer| L[Evidence-Based Output]
    end

Loading

---

🧠 Core Intelligence Modules

1. High-Dimensional Semantic Matching (Cosine Similarity)

Traditional parsers fail on context. We utilize Sentence Transformers (all-MiniLM-L6-v2) to encode text into 384-dimensional vectors.

• The Math: We calculate the cosine similarity between the Job Description vector ($V_{JD}$) and Candidate vector ($V_{Cand}$). $$ \text{Score} = \frac{V_{JD} \cdot V_{Cand}}{|V_{JD}| |V_{Cand}|} $$
• Result: Matches candidates who describe the concept of "Scalable Cloud Architecture" even if they don't explicitly say "AWS".

2. Retrieval-Augmented Generation (RAG)

We implemented a FAISS (Facebook AI Similarity Search) index for real-time candidate interrogation.

• Chunking Strategy: Resumes are split into 300-word sliding windows with 50-word overlap to preserve context boundaries.
• Vector Search: When a user asks "Did they use Python in a production environment?", the system retrieves the nearest neighbor vectors to the query.
• Grounded Answer: These vectors are injected into the Gemini context window, forcing the LLM to answer only based on retrieved evidence, eliminating hallucinations.

3. Weighted Ensemble Scoring Algorithm

Final ranking is not arbitrary. It is a deterministic weighted sum of four distinct analytical axes:

• $\alpha$ (40%) Semantic Alignment: Vector space proximity.
• $\beta$ (30%) Hard Skill Overlap: Regex-based exact matching against SKILLS_DB.
• $\gamma$ (20%) Experience Normalization: LLM-extracted years vs. Requirements.
• $\delta$ (10%) Education Hierarchy: Weighted mapping (PhD > Master's > Bachelor's).

---

🛠️ Technical Stack & Engineering Decisions

Domain	Technology	Engineering Rationale
LLM Orchestration	Google Gemini 1.5 Flash	Selected for high throughput and massive context window (1M tokens) for handling verbose CVs.
Vector Database	FAISS-CPU	In-memory vector search optimized for dense vectors; faster than pinecone/chroma for session-based indexing.
Embeddings	Sentence-Transformers	Uses all-MiniLM-L6-v2 for the optimal balance of inference speed and semantic capture quality.
UI/State	Streamlit	Utilizes @st.cache_resource to keep heavy ML models (Spacy/Transformers) in memory, ensuring sub-second reruns.
Environment	DevContainers	Full infrastructure-as-code. Pre-configured with Python 3.11 and system-level dependencies (build-essential) for FAISS.

---

⚡ Deployment & Infrastructure

🐳 The Dockerized Workflow (DevContainer)

This project is built for Codespaces and Remote Containers. Zero "it works on my machine" issues.

1. System Dependencies: The devcontainer.json automatically runs apt update and installs system libraries required for fitz (PyMuPDF) and faiss.
2. Port Forwarding: Auto-configures port 8501 for instant preview.
3. Extension Pack: Pre-loads ms-python and pylance for immediate intellisense.

🚀 Local Setup (Standard)

# 1. Clone the repository
git clone https://github.com/yourusername/ai-semantic-candidate-matcher.git

# 2. Initialize Virtual Environment
python -m venv venv
source venv/bin/activate

# 3. Install the AI Stack
pip install -r requirements.txt

# 4. Download NLP Artifacts
python -m spacy download en_core_web_sm

# 5. Configure Secrets
# Create .streamlit/secrets.toml and add:
# GOOGLE_API_KEY = "your_key_here"

# 6. Launch Application
streamlit run streamlit_app.py

---

📂 Project Directory Structure

ai-semantic-candidate-matcher/
├── .devcontainer/             # 🐳 Infrastructure as Code
│   └── devcontainer.json      # Docker & VS Code Config
├── .streamlit/                # 🔐 App Configuration
│   └── secrets.toml           # API Keys (GitIgnored)
├── src/                       # (Optional) Module separation
├── streamlit_app.py           # 🚀 Main Application Kernel
├── requirements.txt           # 📦 Pinned Production Dependencies
└── README.md                  # 📄 System Documentation

---

🔮 Future Engineering Roadmap

• GraphRAG Integration: Move beyond vector similarity to Knowledge Graph extraction for mapping candidate relationships (e.g., "Worked at Google" -> implies "High Scalability Experience").
• Multi-Modal Analysis: Use Gemini Pro Vision to analyze graphical resume elements (charts, portfolio thumbnails).
• Asynchronous Processing: Implement Celery/Redis for batch processing of 1000+ resumes.

---

View Source Code | Report Bug

_{Engineered with precision. Powered by Mathematics & Transformers.}