Sararin.ai / Selected work
NBO / NRT on Azure Databricks
From synthetic data to a tested model lifecycle.
Business question
How can an offer-decision experiment connect data preparation, model evaluation, and lifecycle evidence before a real business rollout?
Scope and technical contribution
The case connects business decision framing, synthetic data preparation, model experimentation, and lifecycle engineering. The work covers the path from source data to training, model registration and prediction read-back, with separate acceptance criteria for the pipeline and the model.
Engineering approach
A layered data pipeline and governed model lifecycle separate data validity, model behavior, and business acceptance. Azure Databricks and MLflow supported the experiment.
Architecture decisions
Data preparation, model evaluation, registration and inference read-back are inspected separately. This makes it possible to locate a failure and reproduce the relevant evidence without treating a completed training job as end-to-end success.
Trade-off: synthetic evidence before real use
Synthetic data makes bounded experiments and reproducibility possible without depending on access to an operator dataset. It cannot establish real customer response, policy value or production utility. The scope stays explicit when interpreting every result.
Measured evidence
The bounded experiment used 10,000 synthetic observations. Training, model registration, and prediction/probability read-back were exercised. At the evaluated default decision setting, precision, recall and F1 were 0 because no positives were predicted.
Result and engineering judgment
Pipeline execution and lifecycle checks passed within the experiment. Model discrimination remained weak-to-moderate; a working pipeline did not imply a useful business model. The model-quality failure remains part of the published result, so lifecycle success cannot be mistaken for permission to deploy.
Decision supported by the evidence
The experiment supports a claim of an exercised data-to-model lifecycle and identifies a model-evaluation gap. A real rollout would still need representative validation, meaningful decision outcomes and business acceptance. No revenue uplift or production model quality is claimed.
Limitation
SYNTHETIC_EXPERIMENT_ONLY. These observations are not operator behavior, production performance, or authorization to promote the model. Real-world validation and business acceptance remain necessary.