geronimo 0.1.0

ML deployment framework - from local development to production with one command

Geronimo

ML Development Framework — Build, train, and deploy ML models with production-ready infrastructure.

Geronimo is like dbt for ML: a framework developers build around, not just deployment scaffolding.

Features

Category	Capabilities
SDK	DataSource, FeatureSet, Model, Endpoint, BatchPipeline
Artifacts	Versioned storage for models, encoders, transformers (local/S3/MLflow)
Deployment	Terraform, Docker, CI/CD pipelines, real-time + batch
Monitoring	Metrics, drift detection, alerting
Integrations	MLflow, Snowflake, Postgres, SQL Server, MCP

Installation

# Core installation
pip install geronimo

# With optional integrations
pip install geronimo[mlflow]      # MLflow artifact store
pip install geronimo[databases]   # Snowflake, Postgres, SQL Server
pip install geronimo[all]         # Everything

Or from source:

git clone https://github.com/your-org/geronimo.git
cd geronimo && uv sync

Quick Start

Option A: New Project

# Initialize a new ML project
geronimo init --name credit-risk-model

cd credit-risk-model
uv sync
uv run start

Option B: Import Existing Project

cd /path/to/existing-project
geronimo import .

This generates:

• geronimo.yaml — Deployment configuration
• geronimo_sdk/ — SDK wrappers with TODO tags for manual config

SDK Overview

Data Layer

from geronimo.data import DataSource, Query

training_data = DataSource(
    name="features",
    source="snowflake",
    query=Query.from_file("queries/training.sql"),
)

df = training_data.load(start_date="2024-01-01")

Feature Engineering

from geronimo.features import FeatureSet, Feature
from sklearn.preprocessing import StandardScaler, OneHotEncoder

class CustomerFeatures(FeatureSet):
    age = Feature(dtype="numeric", transformer=StandardScaler())
    segment = Feature(dtype="categorical", encoder=OneHotEncoder())

# Training: fit + transform
X = features.fit_transform(train_df)

# Production: transform only (uses fitted encoders)
X = features.transform(prod_df)

Model Definition

from geronimo.models import Model, HyperParams

class CreditRiskModel(Model):
    name = "credit-risk"
    version = "1.2.0"
    features = CustomerFeatures()

    def train(self, X, y, params: HyperParams):
        self.estimator = XGBClassifier(**params.to_dict())
        self.estimator.fit(X, y)

    def predict(self, X):
        return self.estimator.predict_proba(X)

Artifact Storage

from geronimo.artifacts import ArtifactStore

# Save during training
store = ArtifactStore(project="credit-risk", version="1.2.0")
store.save("model", model.estimator)
store.save("encoder", features.encoder)

# Load in production
store = ArtifactStore.load(project="credit-risk", version="1.2.0")
model = store.get("model")

With MLflow (pip install geronimo[mlflow]):

from geronimo.artifacts import MLflowArtifactStore

store = MLflowArtifactStore(project="credit-risk", version="1.2.0")
store.save("model", model)
store.log_metrics({"auc": 0.95})

Production Endpoints

from geronimo.serving import Endpoint

class PredictEndpoint(Endpoint):
    model_class = CreditRiskModel

    def preprocess(self, request):
        return self.model.features.transform(request["data"])

    def postprocess(self, prediction):
        return {"score": float(prediction[0])}

Batch Pipelines

from geronimo.batch import BatchPipeline, Schedule

class DailyScoringPipeline(BatchPipeline):
    model_class = CreditRiskModel
    schedule = Schedule.daily(hour=6)

    def run(self):
        data = self.model.features.data_source.load()
        predictions = self.model.predict(data)
        self.save_results(predictions)

CLI Commands

Command	Description
geronimo init	Create new ML project
geronimo import	Import existing project with SDK wrappers
geronimo generate all	Generate Terraform, Docker, CI/CD
geronimo generate batch	Generate Metaflow/Airflow pipelines
geronimo validate	Validate configuration
geronimo monitor capture-reference	Capture baseline for drift detection
geronimo monitor detect-drift	Compare current data to baseline

Configuration

# geronimo.yaml
project:
  name: credit-risk-model
  version: "1.2.0"

artifacts:
  store: s3://ml-artifacts/

batch:
  enabled: true
  backend: step-functions
  jobs:
    - name: daily_scoring
      schedule: "0 6 * * *"

monitoring:
  drift_detection:
    enabled: true
    sampling_rate: 0.05
    window_days: 7

Documentation

• Getting Started
• Batch Jobs
• Drift Detection
• MCP Integration

License

MIT