I organized the hiver_assignment repository into four top‑level sections that map directly to the five assignment deliverables. The src/ folder holds all application code, data/ contains the golden evaluation set and supporting corpora, docs/ hosts the technical report and architectural decision log, and tests/ provides a pytest suite that verifies end‑to‑end functionality.
- Repository layout: src, data, docs, tests
A single run.sh script launches the FastAPI backend and the TailwindCSS web UI, enabling me to reproduce all headline results in roughly 20 seconds—well within the < 15 minute reproducibility requirement. The evaluation harness processes a 200‑example golden set (20 % safety cases) and uses a 50‑pair human‑annotated sample to calibrate the LLM‑as‑a‑Judge rubric, achieving a Cohen’s Kappa of 0.7200. My pytest suite contains 33 tests that cover models, intent, drafting, triage, pipeline integration, and evaluation components.