Billed vs ordered quantities
Compare medication quantities to surface shortages, overages, and records that need operational review.
← Selected work
Applied ML platform · synthetic data only
Pharmacy Reconciliation & Prescription Renewal ML Platform
An end-to-end Applied ML system for leakage-safe prescription-renewal prediction, operational follow-up workflows, and containerized deployment using entirely synthetic pharmacy data.
From reconciliation workflow to an operational ML platform.
The project began as a pharmacy billing-vs-ordering reconciliation workflow and evolved into an end-to-end Applied ML system for prescription-renewal prediction and operational follow-up.
Two connected operational questions.
Zero-refill prescriptions
Identify prescriptions that may need renewal review as expected medication supply approaches exhaustion.
System role: prioritize operational follow-up for human review—not provide clinical guidance or automatically close cases.
Designed around time, leakage prevention, and a locked Test.
- 01
Deterministic longitudinal synthetic data
- 02
Point-in-time prediction design
- 03
Leakage prevention
- 04
Chronological Train / Validation / Test split
- 05
16-feature final model input
- 06
Train-only preprocessing
- 07
Time-series cross-validation
- 08
Logistic Regression / Random Forest / XGBoost / LightGBM comparison
- 09
Locked final Logistic Regression before Test
- 10
Threshold = 0.50
- 11
Test consumed once
Recall-first results, without real-world claims.
Interpretation: high recall means the model caught most synthetic renewal cases, with the trade-off of additional false-positive review workload. This is not clinical or real-world performance.
Built beyond the model artifact.
MLflow experiment tracking and locked model artifact
FastAPI single and batch inference
PostgreSQL prediction history
Follow-up case workflow
Activities and explicit case resolution
Alembic migrations
Docker and Docker Compose
Persistent PostgreSQL volume
Health and readiness endpoints
Request IDs and privacy-safe logs
139 automated tests
GitHub Actions CI
Validated locally, with the future path kept separate.
Validated local system
- 01Synthetic dataFixed-seed longitudinal histories
- 02Feature engineeringPoint-in-time features
- 03Leakage-safe temporal MLChronological validation
- 04Locked Logistic RegressionThreshold 0.50
- 05MLflowLocked model artifact
- 06FastAPISingle and batch inference
- 07PostgreSQLPredictions and follow-up cases
- 08Docker ComposeValidated local system
Future architecture — not deployed
AWS components shown in the diagram are planning only and are not represented as deployed technology.
Constraints made explicit.
Chronology first
Used chronological splitting instead of a random split to preserve the temporal prediction setting.
Locked before Test
Selected the model and 0.50 threshold before consuming Test, with no Test-based retuning.
Synthetic by design
All longitudinal pharmacy records and reported model results are deterministic and synthetic.
Prefer interpretable
Selected simpler Logistic Regression after comparison with more complex tree-based models.
Human resolution
Cases require explicit operational resolution rather than automatic closure from model output.
What this work does not claim.
Synthetic data only
The system has not been evaluated on real pharmacy data.
Not a clinical model
Predictions support a simulated operational review workflow, not clinical decisions.
Not production deployed
The validated system runs locally with Docker Compose.
AWS is planning only
The AWS path is future architecture and has not been deployed.
No real-world claim
The reported metrics do not represent real pharmacy performance.
Technical implementation
Explore the complete system.
The repository contains the synthetic-data pipeline, trained model workflow, API, migrations, tests, operational case handling, and Docker Compose setup.
View repository ↗