scoreflow 0.1.0

Production-ready automated ML scoring engine.

ScoreFlow

ScoreFlow is a production-ready automated machine learning scoring engine. It takes raw tabular data, trains and evaluates multiple classifiers, selects the best model by AUC and stability, and exposes it as a REST API or exportable batch pipeline — all from a single Python package.

---

Table of Contents

• Features
• Architecture
• Package Layout
• Installation
• Quickstart
• Modules In Depth
  • Data Ingestion
  • Model Training
  • Evaluation & Selection
  • Scoring API
  • Exportable Pipeline
  • Logging
• CLI Reference
• REST API Reference
• Configuration
• Testing
• Project Roadmap

---

Features

Capability	Details
Multi-model training	Logistic Regression, Random Forest, LightGBM (pluggable interface)
Rigorous evaluation	AUC-ROC, KS statistic, lift at deciles, PSI stability
Automatic selection	Best model chosen by AUC then KS as a tie-breaker
REST API	FastAPI app with /health and /score endpoints
Batch pipeline	Export any trained model to a self-contained folder + run it on CSV/Parquet
Structured logging	JSON-line logs with request IDs, model version, and timing
Reproducibility	Fixed random seeds, versioned artifacts with timestamps
Full test suite	50 tests — unit, API, pipeline, and end-to-end integration

---

Architecture

Raw CSV/Parquet
      │
      ▼
┌─────────────┐    schema, splits
│  Data layer │ ──────────────────────────────────────────────┐
└─────────────┘                                               │
      │                                                       │
      ▼                                                       ▼
┌──────────────┐  ModelArtifact   ┌────────────────┐   ┌───────────────┐
│ Model Trainer│ ────────────────▶│   Evaluator    │──▶│ Model Selector│
└──────────────┘                  └────────────────┘   └───────┬───────┘
                                                               │ best artifact
                                  ┌────────────────┐           │
                                  │  Scoring API   │◀──────────┤
                                  │   (FastAPI)    │           │
                                  └────────────────┘           │
                                                               ▼
                                                    ┌──────────────────┐
                                                    │ Pipeline Exporter │
                                                    │  (offline runner) │
                                                    └──────────────────┘

---

Package Layout

ScoreFlow/
├── scoreflow/
│   ├── config.py               # Paths, seeds, global defaults
│   ├── logging_config.py       # Structured JSON logging setup
│   ├── __main__.py             # CLI: `python -m scoreflow`
│   ├── data/
│   │   ├── schema.py           # DatasetSchema — feature/target definitions
│   │   ├── loaders.py          # CSV/Parquet loaders with validation
│   │   └── splitters.py        # Train / val / test split logic
│   ├── models/
│   │   ├── base.py             # Abstract BaseModel + ModelArtifact dataclass
│   │   ├── logistic.py         # Logistic Regression wrapper
│   │   ├── random_forest.py    # Random Forest wrapper
│   │   ├── lightgbm_model.py   # LightGBM wrapper
│   │   ├── trainer.py          # Trainer orchestrator
│   │   └── registry.py        # Hyperparameter registry
│   ├── evaluation/
│   │   ├── metrics.py          # AUC, KS, lift, PSI
│   │   ├── evaluator.py        # Evaluator — runs all metrics, returns report
│   │   └── selection.py        # select_best_model(), persist_evaluation_report()
│   ├── api/
│   │   └── app.py              # FastAPI app factory (create_app)
│   └── pipeline/
│       ├── exporter.py         # export_pipeline() — serialize artifact to folder
│       └── runner.py           # run_pipeline() — load export, score a file
├── tests/
│   ├── test_config.py
│   ├── test_data.py
│   ├── test_models.py
│   ├── test_evaluation.py
│   ├── test_api.py
│   ├── test_pipeline.py
│   └── test_integration.py     # Full end-to-end flow
├── pyproject.toml
├── Makefile
└── ROADMAP.md

---

Installation

Requirements: Python ≥ 3.11

# 1. Clone the repository
git clone https://github.com/your-org/scoreflow.git
cd scoreflow

# 2. Install with development dependencies
pip install -e ".[dev]"

Core dependencies installed automatically: pandas, scikit-learn, numpy, lightgbm, xgboost, pyarrow, scipy, fastapi, uvicorn, joblib.

---

Quickstart

1 — Train models on your data

import pandas as pd
from scoreflow.data.schema import DatasetSchema
from scoreflow.data.splitters import random_split
from scoreflow.models.trainer import Trainer

df = pd.read_csv("data/my_dataset.csv")

schema = DatasetSchema(target="default")   # name of your binary target column
schema.resolve(df)                         # auto-detect feature columns

splits = random_split(df, target="default", test_size=0.2, val_size=0.1)

trainer = Trainer(schema)                  # trains LR, RF, LightGBM by default
artifacts = trainer.train(splits)
trainer.save_artifacts(artifacts, "models/")

2 — Evaluate and select the best model

from scoreflow.evaluation import Evaluator, select_best_model, persist_evaluation_report

evaluator = Evaluator(schema)
reports = [evaluator.evaluate(a, splits) for a in artifacts]

best = select_best_model(reports, split="val")
print(f"Best model: {best.model_name}  AUC={best.metrics['val']['auc']:.4f}")

persist_evaluation_report(best, "reports/")

3 — Start the scoring API

python -m scoreflow serve --model-dir models/

# Health check
curl http://localhost:8000/health

# Score a record
curl -X POST http://localhost:8000/score \
  -H "Content-Type: application/json" \
  -d '{"records": [{"feature_a": 1.2, "feature_b": 0.5}]}'

4 — Export and run as a batch pipeline

# Export the best model to a portable folder
python -c "
from scoreflow.pipeline.exporter import export_pipeline
from scoreflow.models.base import ModelArtifact
a = ModelArtifact.load('models/logistic_regression')
from scoreflow.data.schema import DatasetSchema
import json, pathlib
schema_data = json.loads((pathlib.Path('models/logistic_regression/schema.json')).read_text())
# (or pass schema from training)
"

# Or use the CLI runner on an exported pipeline
python -m scoreflow run \
  --pipeline exports/logistic_regression \
  --input new_data.csv \
  --output scores.csv

---

Modules In Depth

Data Ingestion

scoreflow/data/schema.py — DatasetSchema

Holds the target column name, optional ID and date columns, and the resolved list of numeric feature columns. Call schema.resolve(df) once to auto-detect features.

schema = DatasetSchema(
    target="default",
    id_col="customer_id",      # optional, excluded from features
    date_col="application_date" # optional, excluded from features
)
schema.resolve(df)
print(schema.features)  # ['age', 'income', 'debt_ratio', ...]

scoreflow/data/loaders.py

from scoreflow.data.loaders import load_dataset

df = load_dataset("data/train.csv")          # CSV
df = load_dataset("data/train.parquet")      # Parquet

Validates: missing rates, data types, required columns.

scoreflow/data/splitters.py

from scoreflow.data.splitters import random_split, time_split

splits = random_split(df, target="default", test_size=0.2, val_size=0.1, seed=42)
# splits.train / splits.val / splits.test

---

Model Training

All models implement a common BaseModel interface: fit(X, y), predict_proba(X), get_params().

Model	Class	Notes
Logistic Regression	LogisticModel	L2 regularised, fast baseline
Random Forest	RandomForestModel	200 trees, bagging ensemble
LightGBM	LightGBMModel	Gradient boosting, high accuracy

Trainer orchestrator

from scoreflow.models.trainer import Trainer

# Train specific models only
trainer = Trainer(schema, model_names=["logistic_regression", "lightgbm"])
artifacts = trainer.train(splits)          # returns List[ModelArtifact]
trainer.save_artifacts(artifacts, "models/")

Each artifact is saved as a sub-directory:

models/
├── logistic_regression/
│   ├── model.joblib       # serialised model
│   └── metadata.json      # name, timestamp, params
├── random_forest/
│   └── ...

---

Evaluation & Selection

Metrics computed (on val and test splits):

Metric	Description
auc	Area Under the ROC Curve
ks	Kolmogorov–Smirnov statistic
lift_10	Lift at top 10% of scores
lift_20	Lift at top 20% of scores
psi	Population Stability Index (train vs val)

from scoreflow.evaluation import Evaluator, select_best_model

evaluator = Evaluator(schema)
reports = [evaluator.evaluate(artifact, splits) for artifact in artifacts]

# Primary sort: AUC ↓, tie-break: KS ↓
best = select_best_model(reports, split="val")

Each report is saved as reports/<model_name>_<timestamp>.json.

---

Scoring API

Built with FastAPI. The app factory pattern (create_app()) isolates state between instances, making parallel test execution safe.

Startup behaviour: At startup, the server scans the --model-dir directory and auto-loads the artifact with the newest timestamp. Override with the SCOREFLOW_MODEL_NAME env var.

Endpoints:

Method	Path	Description
GET	/health	Returns model name + version, or 503 if no model loaded
POST	/score	Scores one or many records, returns probability scores 0–1
GET	/docs	Auto-generated interactive Swagger UI

Example response from /score:

{
  "scores": [0.823, 0.041, 0.671],
  "model_name": "lightgbm",
  "model_version": "2024-11-01T10:32:11"
}

Error codes: 422 for invalid/empty payload, 503 if no model is loaded.

---

Exportable Pipeline

Export a trained model to a portable self-contained folder, then run it anywhere — no full ScoreFlow install needed for scoring.

from scoreflow.pipeline.exporter import export_pipeline

pipeline_dir = export_pipeline(
    artifact,
    schema,
    output_dir="exports/"
)
# exports/logistic_regression/
#   ├── model.joblib
#   └── schema.json

Run the exported pipeline on new data:

from scoreflow.pipeline.runner import run_pipeline

scored_df = run_pipeline(
    pipeline_dir="exports/logistic_regression",
    input_path="new_customers.csv",
    output_path="scores.csv"
)
# scored_df has all original columns + a "score" column (0–1 probability)

---

Logging

ScoreFlow uses Python's standard logging module with an optional JSON-line formatter for production.

from scoreflow.logging_config import configure_logging

configure_logging(json_format=True)   # structured JSON to stdout
configure_logging(json_format=False)  # human-readable (default)

JSON log line example:

{"timestamp": "2024-11-01T10:32:11Z", "level": "INFO", "logger": "scoreflow.api.app", "message": "Auto-selected model artifact: lightgbm"}

---

CLI Reference

# Start the REST API server
python -m scoreflow serve \
  --model-dir models/          # directory of model artifacts (default: models/)
  --model-name lightgbm        # load a specific model (optional)
  --host 0.0.0.0               # bind host (default: 0.0.0.0)
  --port 8000                  # bind port (default: 8000)

# Score a CSV/Parquet file using an exported pipeline
python -m scoreflow run \
  --pipeline exports/lightgbm  # exported pipeline directory
  --input new_data.csv         # input file (CSV or Parquet)
  --output scores.csv          # output file path

The scoreflow command is also available if installed via pip install -e ..

---

REST API Reference

GET /health

Returns the status of the service and the currently loaded model.

Response 200:

{
  "status": "ok",
  "model": "lightgbm",
  "version": "2024-11-01T10:32:11"
}

Response 503 — no model loaded.

---

POST /score

Score one or more records. Each record is a flat JSON object of feature_name → value pairs matching the training schema.

Request body:

{
  "records": [
    {"age": 35, "income": 52000, "debt_ratio": 0.42},
    {"age": 28, "income": 31000, "debt_ratio": 0.65}
  ]
}

Response 200:

{
  "scores": [0.134, 0.812],
  "model_name": "lightgbm",
  "model_version": "2024-11-01T10:32:11"
}

Response 422 — empty records list or invalid feature values.
Response 503 — no model loaded.

---

Configuration

ScoreFlow is configured via environment variables (all optional):

Variable	Default	Description
SCOREFLOW_DATA_DIR	./data	Root directory for raw data files
SCOREFLOW_MODELS_DIR	./models	Root directory for saved model artifacts
SCOREFLOW_REPORTS_DIR	./reports	Root directory for evaluation reports
SCOREFLOW_MODEL_DIR	./models	Model directory used by the API at startup
SCOREFLOW_MODEL_NAME	(auto)	Force-load a specific model sub-directory

Global training defaults live in scoreflow/config.py:

Setting	Value	Description
random_state	42	Fixed seed for reproducibility
test_size	0.2	Fraction of data held out for test
cv_folds	5	Cross-validation folds
primary_metric	roc_auc	Metric used for model selection

---

Testing

# Run the full test suite (50 tests)
python -m pytest tests/ -v

# Run with code coverage
python -m pytest tests/ --cov=scoreflow --cov-report=html
open htmlcov/index.html

# Lint
python -m ruff check scoreflow tests

# Type check
python -m mypy scoreflow

Test coverage by module:

Test file	Tests	Covers
test_config.py	2	Config defaults and path resolution
test_data.py	10	Loaders, schema, splitters, validation
test_models.py	6	All classifiers, trainer, save/load
test_evaluation.py	6	AUC, KS, lift, PSI, evaluator, selection
test_api.py	6	FastAPI health + score endpoints
test_pipeline.py	4	Export structure + batch runner
test_integration.py	1	Full end-to-end: data → train → eval → select → export → run

---

Project Roadmap

See ROADMAP.md for the full phased build plan.

Completed phases:

• ✅ Phase 1 — Foundation & project setup
• ✅ Phase 2 — Data ingestion (loaders, schema, splits, validation)
• ✅ Phase 3 — Model training & selection (LR, RF, LightGBM + trainer + registry)
• ✅ Phase 4 — Evaluation (AUC, KS, lift, PSI, model selection, persisted reports)
• ✅ Phase 5 — Scoring API (FastAPI, /health, /score, model auto-loading)
• ✅ Phase 6 — Exportable pipeline (exporter + batch runner + CLI)
• ✅ Phase 7 — Production hardening (structured JSON logging, integration tests)

Upcoming:

• 🔲 Dockerfile for containerised deployment
• 🔲 Prometheus metrics (request count, latency, error rate)
• 🔲 GitHub Actions CI (lint + test on every push)

---

License

MIT — see LICENSE.