Multimodal E-Commerce Intelligence Prototype
Python · FastAPI · Streamlit · PySpark · PyTorch · LangChain
Overview
A portfolio prototype exploring how several e-commerce analysis components can be exposed through an API and dashboard: a product-description LLM agent, image classification, a PySpark ETL example, and a sales-impact demonstration endpoint.
Components
- Product text analysis: LangChain and an OpenAI chat model return a Pydantic-structured product category, extracted features, a suggested price range, and confidence score. A deterministic mock result is returned when no API key is configured; this is a demo fallback, not an evaluated production model.
- Image classification: A pretrained MobileNetV3-Small model from torchvision returns an ImageNet classification for an uploaded image. It is not trained specifically to identify e-commerce products or manufacturing defects.
- ETL example: PySpark joins order and product data, calculates derived sales values, aggregates category metrics, and writes enriched data to Parquet. The included sample data is generated synthetically.
- Sales-impact endpoint: Returns a formula-based estimate from the supplied discount and rating values. This endpoint is illustrative and is not a validated causal-inference model.
- User interface: A Streamlit dashboard calls the API to demonstrate the available workflows.
Technologies
Python, FastAPI, Pydantic, Uvicorn, Streamlit, Requests, LangChain, LangChain-OpenAI, OpenAI API, PySpark, Pandas, NumPy, PyArrow, Parquet, PyTorch, torchvision, Pillow, Docker, GitHub Actions, pytest.
Scope and Limitations
This is a prototype, not a production-ready analytical platform. The included ETL data is synthetic, the image model is general-purpose, and the sales-impact formula is not a causal estimate supported by an experiment or validated model. The components demonstrate integration patterns and require domain-specific data, evaluation, and operational safeguards before production use.