Pipeline Architecture

This application evaluates clinical trial protocol files by combining low-latency in-memory PDF parsing with LLM zero-shot classification aligned with standard CDISC therapeutic areas.

1. PyMuPDF Metric Sampling

Upon upload, PyMuPDF opens the PDF in memory to inspect structural features (page count, drawing paths, vector table structures) and extracts high-signal executive text from the first 5 pages in under 50ms.

2. Zero-Shot Llama 3.3 Reasoning

The extracted text is routed to Llama 3.3 70B via Together AI to identify biomarkers, primary endpoints, and disease terminology, mapping the document to 1 of 9 CDISC fields (Oncology, Cardiology, Neurology, etc.).

3. Structured JSON Response

The output enforces strict JSON return formats containing match confidence, secondary field overlaps, identified biomarkers, and explicit step-by-step pipeline rationale.