Python

Credit Risk Scorer

End-to-end ML pipeline for predicting loan payback, featuring data prep, model training, MLflow tracking, FastAPI deployment, monitoring, and orchestration.

Source Code Live Demo

Credit Risk Scorer

Credit Risk Scorer trains an XGBoost classifier on LendingClub loan data to predict repayment vs default. Models export to ONNX and register on DagsHub MLflow with experiment tracking, artifact storage, and champion promotion. A FastAPI inference service loads the champion model at startup, serves a scoring UI, and exposes prediction, health, and results endpoints. The pipeline runs via Docker Compose locally or deploys to AWS Lambda as a container image.

What it does

Dataset

The raw data is the LendingClub Loan Data publicly available on Kaggle:

LendingClub Accepted & Rejected Loans 2007–2018 Q4

Download both files and place them in dataset/, then run the ETL to produce the cleaned training file:

PYTHONPATH=. python pipeline/etl.py

Data Versioning

dvc pull          # download data
dvc repro         # re-run ETL if config changed
dvc push          # upload new data to remote

How to Run

Set your DagsHub token before running anything:

export DAGSHUB_USER_TOKEN=<your_token>

With Docker (recommended)

# Train (logs to DagsHub MLflow)
docker compose run --rm train

# Serve the production model
docker compose up inference

Without Docker

Step 1 — ETL (only needed when regenerating dataset/train.csv):

PYTHONPATH=. python pipeline/etl.py

Step 2 — Baselines (majority-class + logistic regression, logged to MLflow):

PYTHONPATH=. python services/train/baselines.py

Step 3 — XGBoost training (trains default XGBoost, exports ONNX, registers + promotes the model):

PYTHONPATH=. python services/train/train.py

Optional — PyTorch training (experimental, not promoted to champion):

pip install -e ".[pytorch]"
PYTHONPATH=. python services/train/pytorch_train.py

Step 4 — Inference API:

PYTHONPATH=. uvicorn services.inference.app:app --reload

Then open http://localhost:8000 for the built-in scoring UI.

Exploratory Analysis

PYTHONPATH=. python notebooks/feature_analysis.py
PYTHONPATH=. python notebooks/profile_dataset.py   # generates docs/train_profile.html

Load Testing

locust -f tests/load_test.py --host http://localhost:8000

EC2 Deployment

The inference API can run on a persistent EC2 instance as a Docker container (uvicorn via services/inference/inference.dockerfile). Deployments are triggered manually through .github/workflows/deploy.yaml (workflow_dispatch only — auto-deploy on push is disabled).

The workflow builds the inference image, pushes it to private ECR, SSHes into the instance, pulls the image, and restarts the loanpayback container on port 80.

GitHub secrets:

Secret	Example
`AWS_ACCESS_KEY_ID`	IAM access key for CI
`AWS_SECRET_ACCESS_KEY`	IAM secret key
`AWS_REGION`	`us-east-1`
`ECR_REGISTRY`	`031749757344.dkr.ecr.us-east-1.amazonaws.com`
`ECR_REPOSITORY`	`credit-models`
`EC2_HOST`	Public IP or DNS of the instance
`EC2_USER`	`ec2-user` (Amazon Linux) or `ubuntu`
`EC2_SSH_KEY`	Private key contents for SSH

One-time EC2 setup:

# On the instance: install Docker + AWS CLI (Amazon Linux 2023 example)
sudo dnf install -y docker awscli
sudo systemctl enable --now docker
sudo usermod -aG docker ec2-user

# Create ECR repo (if not already created)
aws ecr create-repository --repository-name credit-models

# Security group: allow inbound TCP 80 (and 22 for SSH)

Store DAGSHUB_USER_TOKEN and ADMIN on the instance (e.g. in ~/.env) — the container needs them at runtime to reach MLflow and protect /reload.

Manual deploy (from repo root):

source .venv/bin/activate
set -a && source .env && set +a

aws ecr get-login-password --region us-east-1 \
  | docker login --username AWS --password-stdin 031749757344.dkr.ecr.us-east-1.amazonaws.com

docker build -f services/inference/inference.dockerfile -t credit-models .
docker tag credit-models:latest \
  031749757344.dkr.ecr.us-east-1.amazonaws.com/credit-models:latest
docker push 031749757344.dkr.ecr.us-east-1.amazonaws.com/credit-models:latest

# On EC2 (or via SSH)
docker stop loanpayback || true && docker rm -f loanpayback || true
docker pull 031749757344.dkr.ecr.us-east-1.amazonaws.com/credit-models:latest
docker run -d \
  -p 80:8000 \
  --name loanpayback \
  --env-file ~/.env \
  031749757344.dkr.ecr.us-east-1.amazonaws.com/credit-models:latest

The API is then available at http://<EC2_HOST>/. Unlike Lambda, the container stays warm — no cold starts after the initial model load.

Note: EC2 and Lambda currently push to the same ECR tag (credit-models:latest) but use different Dockerfiles. Redeploying one overwrites the image the other expects — use separate tags (e.g. :ec2 and :lambda) if running both in production.

Lambda Deployment

The inference API runs on AWS Lambda as a container image. Pushes to main trigger .github/workflows/deploy-lambda.yaml, which builds and pushes to private ECR (Lambda does not support ECR Public for function images).

GitHub secrets (in addition to existing AWS / EC2 creds):

Secret	Example
`LAMBDA_FUNCTION_NAME`	`credit-risk-inference`
`ECR_REGISTRY`	`031749757344.dkr.ecr.us-east-1.amazonaws.com`
`ECR_REPOSITORY`	`credit-models`

Manual deploy (from repo root):

source .venv/bin/activate
set -a && source .env && set +a

aws ecr get-login-password --region us-east-1 \
  | docker login --username AWS --password-stdin 031749757344.dkr.ecr.us-east-1.amazonaws.com

docker build --provenance=false --sbom=false \
  -f services/inference/lambda.dockerfile -t credit-models .

docker tag credit-models:latest \
  031749757344.dkr.ecr.us-east-1.amazonaws.com/credit-models:latest

docker push 031749757344.dkr.ecr.us-east-1.amazonaws.com/credit-models:latest

aws lambda update-function-code \
  --function-name credit-risk-inference \
  --image-uri 031749757344.dkr.ecr.us-east-1.amazonaws.com/credit-models:latest

aws lambda wait function-updated-v2 --function-name credit-risk-inference

One-time AWS setup:

# IAM role for Lambda
aws iam create-role \
  --role-name lambda-inference-role \
  --assume-role-policy-document '{
    "Version": "2012-10-17",
    "Statement": [{
      "Effect": "Allow",
      "Principal": {"Service": "lambda.amazonaws.com"},
      "Action": "sts:AssumeRole"
    }]
  }'

aws iam attach-role-policy \
  --role-name lambda-inference-role \
  --policy-arn arn:aws:iam::aws:policy/service-role/AWSLambdaBasicExecutionRole

aws iam attach-role-policy \
  --role-name lambda-inference-role \
  --policy-arn arn:aws:iam::aws:policy/AmazonEC2ContainerRegistryReadOnly

# Private ECR repo (if not already created)
aws ecr create-repository --repository-name credit-models

# Create Lambda function (secrets via env, not baked into image)
aws lambda create-function \
  --function-name credit-risk-inference \
  --package-type Image \
  --code ImageUri=031749757344.dkr.ecr.us-east-1.amazonaws.com/credit-models:latest \
  --role arn:aws:iam::ACCOUNT_ID:role/lambda-inference-role \
  --timeout 120 \
  --memory-size 3008 \
  --architectures x86_64 \
  --environment "Variables={
    DAGSHUB_REPO_OWNER=pjawale,
    DAGSHUB_REPO_NAME=credit-scorer,
    DAGSHUB_USER_TOKEN=$DAGSHUB_USER_TOKEN,
    ADMIN=$ADMIN
  }"

# Public Function URL
aws lambda create-function-url-config \
  --function-name credit-risk-inference \
  --auth-type NONE \
  --cors '{"AllowOrigins":["*"],"AllowMethods":["*"],"AllowHeaders":["*"]}'

aws lambda add-permission \
  --function-name credit-risk-inference \
  --statement-id FunctionURLAllowPublicAccess \
  --action lambda:InvokeFunctionUrl \
  --principal "*" \
  --function-url-auth-type NONE

aws lambda add-permission \
  --function-name credit-risk-inference \
  --statement-id PublicInvokeFunction \
  --action lambda:InvokeFunction \
  --principal "*"

Expect 30–70s cold starts while the container imports dependencies and loads the ONNX model from MLflow. Warm /predict calls are sub-second.