🎵 VibeStudent: Hybrid LLM & ANN-Powered Music Recommendation Engine

VibeStudent is an enterprise-scale, hybrid music recommendation architecture bridging semantic music intelligence with real-time, low-latency inference. By applying Knowledge Distillation from a high-capacity Large Language Model (Llama 3.2) to a lightweight Artificial Neural Network (ANN), the system maps multi-modal audio signals, artist semantics, and metadata into a high-dimensional vector space for real-time similarity search over massive datasets.

---

🚀 Key Architectural Highlights

• Knowledge Distillation (Teacher-Student Pipeline): Translates deep contextual and cultural music understanding from an on-premise LLM (Llama 3.2 via Ollama) into a fast, 256-dimensional neural network embedding layer (VibeStudent).
• Multi-Modal Feature Space: Ingests a 75-dimensional input vector combining:
  • Audio Features (9–11D): Danceability, energy, tempo, valence, acousticness, etc.
  • Artist Latent Space (64D): Trainable artist embeddings capturing genre clustering and musical style.
  • Metadata (2D): Normalized release year and track popularity.
• Vector Indexing & Sub-Millisecond Search: Integrates FAISS (Facebook AI Similarity Search) with $O(\log N)$ search complexity to query over 1.16 million indexed tracks with minimal RAM footprint (~1.1 GB).
• Production Stability & Data Drift Mitigation:
  • Frozen Global Scalers: Prevents vector space drift when ingesting streaming tracks.
  • Batch Normalization: Guarantees latent stability across varying feature distributions.
• Continuous Self-Improvement via Reinforcement Learning (RL): Adapts to live implicit user feedback (play completion vs. early skip) using policy gradient updates constrained by KL-Divergence, Experience Replay, and a Target Network to avoid catastrophic forgetting.
• Scalable Data Ingestion (Lambda Architecture): Ready for distributed stream ingestion via Apache Kafka and Spark Streaming, separating the immutable batch layer (SQL/Parquet) from the real-time speed layer.

---

🧠 Neural Architecture Details

Input Layer (75 Dimensions)
  ├── 9 Audio Features (Acoustic / Psychoacoustic)
  ├── 64D Trainable Artist Embedding Lookup
  └── 2 Metadata Parameters (Year, Popularity)
         │
         ▼
Hidden Bottleneck / Representation Layer (256 Neurons, ReLU, BatchNorm)
  └── [Extracts 256D "Vibe" Coordinate Vector -> FAISS Index]
         │
         ▼
Output Layer (2000 Neurons, Softmax)
  └── Soft Genre Distribution (Confidence Score for Fallback & Auditing)

---

📊 Evaluation & Validation Metrics

The engine measures both retrieval relevance and ranking quality against distilled ground-truth benchmarks:

• Precision@K: Evaluates the concentration of relevant items within the top-$K$ recommendations.
• nDCG (Normalized Discounted Cumulative Gain): Penalizes relevant recommendations positioned lower in the ranked list to guarantee peak relevance in top ranks.
• Confidence Gating: Utilizes Softmax probability distributions. Low-confidence inferences trigger an automated fallback to the Teacher LLM for supervision and continuous model calibration.

---

🛠️ Tech Stack & Dependencies

• Core: Python 3.10+, PyTorch, NumPy, Pandas, Scikit-learn
• Vector Engine: FAISS (CPU/GPU)
• LLM Engine: Llama 3.2 running on-premise (Ollama)
• Data Pipelines: Apache Kafka, Apache Spark Streaming, SQLite / PostgreSQL, Parquet
• Hardware Acceleration: AMD ROCm / CUDA

---

⚙️ Quickstart Guide

1. Clone the Repository

git clone [https://github.com/Golden-Stone27/Spotify_Project.git](https://github.com/Golden-Stone27/Spotify_Project.git)
cd Spotify_Project

2. Environment Setup

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -r requirements.txt

3. Configure Environment Variables

Create a .env file in the root directory:

SPOTIFY_CLIENT_ID="your_spotify_client_id"
SPOTIFY_CLIENT_SECRET="your_spotify_client_secret"
OLLAMA_ENDPOINT="http://localhost:11434"
LLM_MODEL_NAME="llama3.2"
EMBEDDING_DIM=256

4. Run the Pipeline

# Data preparation and feature scaling
python src/data_preprocessing.py

# Model training with Knowledge Distillation
python src/train.py

# Build and query FAISS vector index
python src/indexing.py

---

🗺️ Roadmap

• Initial KNN baseline and exploratory data analysis.
• Multi-modal feature processing (64D artist embeddings + audio descriptors).
• Knowledge distillation pipeline with Llama 3.2.
• FAISS indexing for 1.16M tracks.
• Online Reinforcement Learning pipeline with Experience Replay.
• Distributed stream processing via Apache Kafka and Spark Streaming.
• Automated hyperparameter scheduling using Bayesian Optimization.

---

📄 License

Distributed under the MIT License. See LICENSE for more information.