Automate Drug Inventory and Reordering with a Custom AI System

A multi-location dermatology clinic manages high-value injectables with specific expiration dates and lot numbers. The practice manager spends every Friday afternoon manually counting vials in three separate refrigerators. This manual process is slow and consistently leads to errors, either through overstocking temperature-sensitive drugs or running out of a popular filler.

They tried using their EHR's built-in inventory module, but it only tracks decrements when a service is billed. It has no forecasting, so a sudden increase in demand for Botox leads to a stockout, forcing them to reschedule patients. The module also cannot handle unique lot number tracking for multiple vials of the same drug, creating compliance risks during an audit.

A generic inventory app like Zoho Inventory was even worse. It lacks fields for NDC (National Drug Code) numbers and cannot enforce first-in, first-out usage based on expiration dates. This caused them to use a newer vial before an older one, which then expired, costing them $600. The app also is not designed for healthcare data security, making a connection to patient records impossible.

Syntora would begin by understanding your current drug inventory challenges and existing technical infrastructure. The first technical step involves securely connecting to your practice's EHR, either through its documented API or a secure database connection. We would pull relevant historical data, typically 12-24 months of appointment and billing records, to establish a baseline for medication consumption. This process maps procedure codes to specific drugs and dosages, creating a historical consumption record for all medications. Data extracted would be stored in a secure, scalable database like Supabase Postgres. We have experience building similar data ingestion pipelines for sensitive information, ensuring data integrity and security.

This historical data would then be used to train tailored forecasting models for each medication. For drugs with stable demand, a model like ARIMA can be effective. For seasonal items, such as flu vaccines, a more complex model like LightGBM would be employed, incorporating features such as time of year, patient demographics, and recent appointment volume to improve accuracy. The goal of this modeling is to predict demand for the next 30 days with a high degree of precision, allowing for the proactive adjustment of reorder points. All prediction logic would be developed using Python.

The core reordering logic would be implemented as a FastAPI application. This application would monitor predicted inventory levels against established reorder points. When a threshold is met, the system would generate a purchase order. For suppliers offering modern APIs, such as McKesson, the system would be designed to post orders directly. For suppliers relying on web portals, we would implement automation using tools like Playwright to handle login and form submission. We have built document processing pipelines using Claude API for financial documents, and a similar pattern applies to extracting data or submitting orders via web portals if direct APIs are unavailable. To maintain client control, every generated order would typically require a manual approval step from a practice manager, often facilitated through a secure link delivered via a communication platform like Slack.

The FastAPI service would be containerized using Docker and deployed to a serverless platform such as AWS Lambda. This architecture provides scalability and cost efficiency, with typical operational costs for a clinic often remaining under $30 per month. The system would maintain a comprehensive audit trail in Supabase, logging every action, including predictions, purchase order generations, and approvals. Monitoring would include CloudWatch alerts configured to notify relevant personnel of any order failures or system anomalies. The goal for order placement API response times would be under 300ms for critical operations.