Add CLAUDE.md with project conventions and architecture notes

Author: jgbustos

CLAUDE.md

@@ -0,0 +1,63 @@
+# CLAUDE.md
+
+## Project Overview
+A RESTful API that serves predictions from a trained ML model, built with Python 3 and Flask-RESTX.
+
+## Environment
+- Python 3.13 (venv at `venv/`)
+- Activate venv: `source venv/Scripts/activate` (bash on Windows)
+
+## Quality Gate
+Run before committing:
+```
+powershell.exe -File make.ps1
+```
+Checks: black (formatter) → pylint → mypy. Must score 10.00/10 and zero mypy errors.
+
+## Dependencies
+- `requirements.txt` — pinned runtime deps (Flask, flask-restx, numpy, pandas, loguru, etc.)
+- `requirements-dev.txt` — dev tools (black, pylint, mypy)
+- `tests/requirements.txt` — test deps (pytest, requests, openapi-spec-validator)
+- Update all three when installing new packages in the venv
+
+## Git Workflow
+- Branch: `master`
+- Remote: `https://github.com/jgbustos/ml_rest_api.git`
+- Git user email: `35463898+jgbustos@users.noreply.github.com` (GitHub no-reply, required to push)
+- Run quality gate before every commit
+- Never commit `.claude/` directory
+
+## Key Architecture Decisions
+
+### Logging (loguru)
+- Entry point: `ml_rest_api/logging_setup.py` → `setup_logging()`
+- Called once in `app.py` after `Flask(__name__)`
+- `InterceptHandler` redirects all stdlib `logging` calls to loguru
+- Multi-line log messages are split per line so each gets a loguru header
+- `click.echo` and `click.secho` are monkey-patched to capture Flask/Werkzeug `* startup lines`
+- All modules keep using `getLogger(__name__)` — no need to import loguru directly
+
+### CSRF Protection
+- `CSRFProtect` is bound to the app in `initialize_app()` after `configure_app()`
+- Controlled via `WTF_CSRF_ENABLED` setting (default: `True`)
+
+### Type Annotations
+- Use built-in generics (`dict[str, Any]`, `list[str]`) not `typing.Dict`/`typing.List`
+- flask-restx has incomplete stubs — suppress with `# type: ignore` at import and decorator sites
+- Pylance strictness is high; suppress false positives with `# type: ignore[specific-code]`
+
+### ML Model
+- Model module loaded dynamically via `TrainedModelWrapper` in `wrapper.py`
+- Module name configured via `TRAINED_MODEL_MODULE_NAME` env var or settings
+- `ml_trained_model.py` is the default/example model module
+
+## Docker
+- `Dockerfile` uses `ENTRYPOINT ["python3"]` + `CMD ["ml_rest_api/app.py"]`
+- `docker-compose.yml` `command` must only pass the script path, not `python3` again
+- `.dockerignore` excludes: venv, tests, nginx, docker-compose, CI config, dev requirements
+
+## CI/CD
+- CircleCI config: `.circleci/config.yml`
+- Python image: `cimg/python:3.13`
+- Cache key: `v1-dependencies-py313-...`
+- Requires a CircleCI checkout key (User Key via GitHub OAuth)