2024client work
Kidney Disease Detection
I focus on reproducibility before worrying about model scoresPipeline stats and every run is repeatable
problem
The challenge: a model is only reliable if anyone can trace its data parameters and steps
I make metrics and retrains clear and reproducible
- Every pipeline stage tracked from raw data to model registry
- All experiments logged with MLflow and DVC
- Used AWS EC2 and S3 for speed and trackable artifacts
- Consistent 85% test accuracy
pipeline
decisions
Track the dataset, split, and every stage with DVC rather than git-only metrics.
tradeoffVersioning ceremony on every retraining change.
outcomeA regression is bisectable to a stage, not to a folder.
Train on EC2 with model and dataset artifacts on S3, not bulk-uploaded notebooks.
tradeoffExtra IAM configuration on the training node.
outcomeReproducible pipelines exist without assuming the training environment is literally the analyst's laptop.
results
- test accuracy
- 85%
- stages
- DVC
- raw → register → serve
- artifacts
- S3
- data and model versions
stack
Modeling
- PyTorch
- scikit-learn
MLOps
- MLflow
- DVC
- params.yaml
Cloud
- AWS EC2
- AWS S3
- IAM
Delivery
- GitHub Actions
- Docker