Pharma enterprise customer (literature screening): Pharma Literature Screening Automation case study
Pharmacovigilance literature screening for an Australian operator, built for a pharma enterprise customer. A weekly manual spreadsheet process became audited software that reproduces the operator's Excel tracker to the cell, uses Claude on AWS Bedrock in two tiers and routes uncertain cases to a person. POC.
The Problem
Pharma companies must screen local literature and regulator sources every week for safety information about their products, and prove to a regulator that they did. The operator did this by hand: a reviewer searched each approved term in each source and recorded the result in a spreadsheet. Almost every weekly row is a zero-hit result, and that negative result is the deliverable, so every search has to be shown and every conclusion traceable.
The Solution
Apex36 built the screening software for a pharma enterprise customer. Each weekly run acquires the sources, extracts the text, classifies it and exports the client's own Excel tracker, reproduced to the cell, including the exact free-text wording.
Claude on AWS Bedrock runs in two tiers: a small model for triage and a larger one for the detailed safety review, reached over a private VPC endpoint in the Sydney region. An individual case safety report verdict needs all four regulatory criteria, each backed by a verbatim quote; with fewer, the system downgrades the verdict. Cases the model cannot decide are marked indeterminate and routed to a human reviewer, whose name goes into the tracker.
Features
Excel tracker to the cell
The export matches the operator's own tracker, column for column.
Two-tier Claude on AWS Bedrock
Triage with a small model, detailed review with a larger one, over a private VPC endpoint.
Evidence for every verdict
Each safety criterion needs a verbatim quote or an explicit statement of absence.
Abstain to a human
Uncertain cases go to a reviewer who signs off in the app.
Append-only audit trail
Corrections are new records, and updates and deletes are blocked at the database role level.
Blocked runs are visible
A source that cannot be searched shows as blocked, not as silence.
Results / Impact
by software that reproduces the tracker to the cell.
for every run, query, verdict and sign-off.
with weekly runs scheduled by EventBridge.
stated as such.
FAQ
It is the weekly search of local literature and regulator sources for safety information about a company's products, with a record that proves every source was searched. Most weeks find nothing, and that negative result must still be evidenced.
A safety verdict needs a verbatim quote for each required criterion, or the system downgrades it. Cases the model cannot decide are routed to a human reviewer instead of being forced into a category.
Claude runs on AWS Bedrock in the Sydney region, reached over a private VPC endpoint, and each call's cost is recorded.
No. It is a POC, stated as such.
Ready to build something impactful?
Let's discuss your project and how we can help you ship faster and smarter.
Book a Free Strategy Call