| # LLM Price Prediction Engine | |
| The Price Is Right is an end-to-end machine learning system that predicts the retail price of a product from its text description alone — title, category, brand, and feature details. It covers the full ML lifecycle from data engineering through classical ML, deep learning, and LLM fine-tuning, with rigorous evaluation across all approaches. | |
| ## What It Does | |
| Given a product description, the system predicts its USD retail price. The problem is harder than it sounds — prices span three orders of magnitude ($0.50 to $999) and are influenced by subtle signals like material quality, brand prestige, and feature density. | |
| ## The Approach | |
| The project benchmarks 12 models across five approaches in order of increasing sophistication: a constant baseline, traditional ML (Linear Regression, Random Forest, XGBoost), a custom deep neural network, frontier LLMs zero-shot (Gemini 2.5 Flash), and finally fine-tuned open-source LLMs. | |
| The DNN uses a HashingVectorizer with 5,000 features, 8 residual blocks, and log-normalized price targets — necessary because raw prices span a 2000x range. | |
| The best result comes from fine-tuning Llama 3.2-3B using QLoRA — 4-bit NF4 quantization plus LoRA adapters — which reduces trainable parameters 163x from 3 billion to 18.4 million. This makes it trainable on a free Colab T4 GPU. The fine-tuned model treats price prediction as text completion, generating "Price is $X.00" and extracting the number via regex. | |
| ## Results | |
| Fine-tuned Llama 3.2 achieved $39.85 MAE, outperforming Gemini 2.5 Flash zero-shot ($62.84) and the human benchmark ($87.62) — a 54% improvement over human performance. XGBoost scored $68.23 MAE as the best traditional ML result. | |
| ## Data Pipeline | |
| 400,000+ Amazon product listings are loaded in parallel using ProcessPoolExecutor, cleaned, then summarised by a fast LLM into a standardised 5-field format before fine-tuning. ThreadPoolExecutor handles concurrent LLM API calls during preprocessing. Experiment tracking uses Weights and Biases. | |
| ## Tech Stack | |
| Python, PyTorch, scikit-learn, XGBoost, Llama 3.2, QLoRA, PEFT, TRL, BitsAndBytes, Gemini 2.5 Flash, HuggingFace, Weights and Biases, Google Colab. |