From prototype to product

The notebook works. Now it has to run 10 000 times a day for people who are not you. The difference:

Prototype	Product
cells in order	modules: data.py, model.py, service.py with side-effect-free functions
hard-coded paths and keys	configuration via the environment or a file, secrets outside the code
«it worked when I ran it»	tests plus an eval harness run in CI on every change
print	structured logging with a request id, without personal data
one call	an API with validation, timeouts, error codes, quotas
the best case	degraded modes when the model or the provider is down
«the model»	a versioned artefact with a data hash and eval results

Do it in that order. Tests and evals first — they pay for themselves at the first change.

project/
├── src/svc/
│   ├── config.py      # pydantic-settings: MODEL_PATH, LLM_URL, LOG_LEVEL …
│   ├── data.py        # load/clean — pure functions
│   ├── model.py       # predict(text) -> Out ; loaded once
│   ├── service.py     # FastAPI: /healthz, /v1/predict, error handling
│   └── log.py         # json logging with request_id
├── tests/             # pytest: data, model, API (httpx.AsyncClient(app=app))
├── evals/cases.jsonl  # the harness; the threshold in CI
├── Dockerfile
├── requirements.txt   # pinned versions
└── .github/workflows/ci.yml   # lint → test → eval → build

# config.py
from pydantic_settings import BaseSettings
class Settings(BaseSettings):
    model_path: str = "artifacts/model-v3.pkl"
    llm_url: str
    llm_timeout_s: float = 20
    daily_budget_sek: float = 100
settings = Settings()   # reads the environment; a missing llm_url → a clear error at start-up, not at the first call