Unsupervised Anomaly Detection Model

I have a collection of unstructured data—mixed text documents and images—and I need an unsupervised learning workflow that reliably flags unusual or suspicious examples. The aim is purely anomaly detection; no labels are available and none can be added, so clustering or classification won’t help here.

Here’s the flow I have in mind:

• Data preparation: consistent preprocessing for both modalities (tokenisation or embeddings for text, feature extraction for images).
• Model development: an unsupervised architecture such as auto-encoder, variational auto-encoder, deep clustering, or another approach you can justify for anomaly detection.
• Evaluation: quantitative metrics (reconstruction error distributions, AUC, or similar) plus a concise report explaining thresholds and decision logic.
• Deliverables: clean, well-commented Python code (ideally PyTorch or TensorFlow/Keras), reproducible environment files, and a short README so I can retrain or fine-tune later.

I will supply a representative sample to start; please keep the design modular so it scales once the full dataset arrives. Let me know any additional dependencies you anticipate, along with an outline timeline for data exploration, model iteration, and final validation.