← back to all projects

Document layout analysis & OCR pipeline

89% F1 on held-out splits

PyTorch · LayoutLMv3 · HuggingFace · FastAPI · Docker · MLflow · W&B

Placeholder: expand this write-up with architecture notes, screenshots, and results.

Fine-tuned LayoutLMv3 on PubLayNet (~360K annotated regions) for multi-class document region segmentation covering text, tables, and figures.

Ran 12+ systematic hyperparameter experiments tracked via MLflow and Weights & Biases, reaching 89% F1 on held-out test splits. Engineered a structured JSON extraction layer on top of model outputs for downstream usability, and deployed the full inference pipeline as a Dockerized FastAPI microservice with MLflow Model Registry for versioned model serving.