This journey was all about repeating a solid development cycle across multiple domain models.
"""
Requirements Definition ➔ API Specification & Design ➔ Infrastructure & Core Implementation ➔ Swagger API Testing ➔ Git PR & Code Review
"""
📌 Stage 4 — User Authentication & Authorization APIs (9 Endpoints)
Design & Review: Converted requirement specifications (REQ-USER-001~009) into detailed API mapping tables. During reviews, I corrected common mistakes like putting hardcoded data into endpoints, swapping Request Body with Response structures, and properly distinguishing Authentication (login status) from Authorization (role permissions).
Implementation: Built core security features using JWT and password hashing (core/security.py). I manually filled in logic for 9 endpoints, catching frequent async bugs like missing await statements.
Testing & Git: Verified full user workflows in Swagger UI (Registration ➔ Login ➔ Auth Header ➔ Profile Fetch). Encountered a Git branching mistake by accidentally setting main as the base branch for a PR, which taught me how to fast-forward sync dev with main.
📌 Stage 5 — Patient & Medical Record APIs with File Upload (8 Endpoints)
Design & Domain Logic: Handled requirements for Patients and Medical Records. Discovered key database behaviors like how ON DELETE CASCADE automatically deletes associated records and X-rays when a patient is removed. Also mapped out multi-part file uploads for X-ray images.
New Concepts Learned: Applied Query Parameters (search/filtering), Path Parameters (/{patient_id}), and Multipart/Form-Data file handling.
Swagger Testing: Successfully tested the end-to-end flow: Creating a patient ➔ Uploading actual X-ray images to the media/xray/ directory ➔ Retrieving medical records with linked media.
📌 Stage 6 — AI Pneumonia Prediction (Full Workflow)
- Model Pipeline & Ensemble Setup
Environment & Setup: Installed PyTorch CPU version (uv add torch torchvision --index ...) and analyzed 10 pre-trained .pth checkpoints (ConvNeXt-Tiny 5-fold + DenseNet121 5-fold).
Domain Verification: Re-verified preprocessing specifications (224x224 Resize, ImageNet Normalization) and the OR-ensemble strategy (predicts pneumonia if either model flags it) from an earlier project report (폐렴Xray분류프로젝트보고서.docx).
Git Strategy & Debugging: Added heavy model weights (worker/models/*.pth) to .gitignore to stay under GitHub's 100MB limit. Debugged key issues, including a loop return bug that loaded only a single fold, and a tuple syntax mistake in Resize((224, 224)).
Inference Test: Tested with actual X-ray images—achieved accurate predictions with a 0.9986 / 0.9888 confidence score in just 0.78 seconds!
- API Design & Requirements Alignment
Requirements Reconciliation: Compared past design documents against current requirements (REQ-PRED-001/002). Identified key updates: expanded access rights from admin/doctor-only to all departments, and made heatmap URLs optional.
Database Precision Update: Updated DB precision for prediction confidence scores from DECIMAL(5,2) to DECIMAL(5,4) and allowed heatmap_url to be nullable.
- API Implementation & Asynchronous Execution
Database Migration: Ran alembic revision --autogenerate and alembic upgrade head to apply schema updates, verifying accuracy directly via DESCRIBE table queries.
Non-blocking AI Inference: Integrated FastAPI’s run_in_threadpool to run heavy, synchronous PyTorch CPU inference on a background thread without blocking the async event loop!
Logic Implementation: Written and refined run_ai_analysis and list_ai_analysis functions after resolving missing await keywords.
- End-to-End Testing & Git Repository Clean-up
Swagger Testing: Verified full user flows: User Registration ➔ Login ➔ POST /ai-analysis (201 Created with is_pneumonia: true, confidence: 0.9986) ➔ Duplicate request caching check (200 OK) ➔ Error handling (404/401/403).
Git Clean-up: Created PR #9 to dev, synchronized dev to main via fast-forward, and cleaned up the repository by deleting 9 merged branches (both local and remote), leaving a clean repository with only main and dev!
