Case study · Pharma

90 days to a reviewed,
auditable extraction pipeline.

A mid-cap pharmaceutical company's regulatory affairs team was manually extracting endpoint data from clinical study reports. We shipped a frontier-model pipeline that cut reviewer time by 73% — without changing their SOPs.

73%
↓ reviewer time
Across 240 study reports in the first quarter.
99.1%
accuracy
Against expert-labeled ground truth set.
0
audit findings
External regulatory review of shipped system.
90
days to prod
From engagement start to first live document.

The problem

Three senior people, reading the same document.

Every clinical study report coming back from a CRO was being read cover-to-cover by a regulatory associate, re-read by a reviewer, and spot-checked by the VP. The team knew it was the wrong shape of work — but the outputs flowed into regulatory submissions, so they couldn't automate something they couldn't audit.

What we built

A pipeline their QA team reviewed as if it were a person.

A frontier-model extraction pipeline with a schema-validated output, expert-in-the-loop review UI, full prompt + response audit trail, and a versioned evaluation set labeled by their own scientists. Every run is logged, attributable, and reviewable.

How we got there

Four stages, 90 days.

Week 1–2

Readiness assessment

Interviewed 9 people across regulatory, QA, and IT. Mapped 14 document-review workflows. Picked endpoint extraction as the pilot: highest volume, cleanest ground truth, bounded downside.

Week 3–8

Measured pilot

Built the eval set — 120 reports labeled by the client's scientists. Shipped three model variants, compared against the human baseline. Settled on Claude 3.5 Sonnet with a structured output schema and a self-review step.

Week 9–12

Production system

Shipped inside their Azure tenant. SSO via Entra, audit trail into their existing SIEM, review UI for the RA team, cost caps, and alerting. Runbooks handed to their ops team.

Ongoing

Governance & scale

Quarterly review with RA leadership, compliance, and QA. Evaluation set expanded to 340 reports. Two more workflows now in the pilot stage.

"
Aiveris didn't sell us an AI platform. They shipped a system we actually operate — with evaluations our scientists trust and an audit trail our compliance team signed off on. That's the part nobody else was willing to do.
VP, Regulatory Affairs · Mid-cap pharmaceutical

Let's talk

Start with a 30-minute call.

We'll walk through one workflow you've been trying to improve, and leave you with a concrete view of what a pilot would look like — whether you hire us or not.

Book a call → Send a note