Pharma’s $100B data problem isn’t the report

My co-founder Margish — his family has manufactured active pharmaceutical ingredients in India for over 30 years. These are the molecules that actually make medicines work.

A few months ago, I’m visiting him in San Diego. His dad is walking me through what’s been keeping his team up at night: the APQR — an annual product quality review. It’s the document that proves to regulators and customers that your drugs are safe. In pharma manufacturing, when a customer trusts you, they stay with you for decades — and that trust lives or dies on how well you can show your process stayed in control.

So I start asking questions. How does this report actually get made? And the answer floored me. It’s manual. Quality teams spend four to eight weeks pulling data from scattered systems — batch records, deviation logs, stability data — just to assemble one report for one product. A mid-size manufacturer has fifty to a hundred products. They’re doing this all year, every year.

My first thought was: why isn’t anyone using AI for this? So I brought in Parsa — my co-founder, machine learning engineer, built AI products at AWS — and we started digging. That’s when we realized: the report isn’t the real problem. It’s a symptom. These companies generate enormous amounts of quality data and they have no way to see what it’s telling them. They’re losing five to thirty percent of revenue — over a hundred billion dollars industry-wide— to quality failures they could have predicted and prevented.

That’s what we’re building. Veriqa doesn’t just draft the report. It turns scattered quality data into intelligence — so manufacturers can see the failures coming before they happen. The report is just one output from the system.

The report is a symptom

If you’ve never worked in pharma manufacturing, an APQR — annual product quality review — is the document that gets sent up the chain once a year for every product you make. The FDA requires it. So does the EMA. ICH Q7 and Q10 codify it. A proper APQR covers every batch produced that year, every deviation, every complaint, every stability trend, every critical process parameter. It’s a regulator’s way of asking, in writing, one question: did this manufacturing process stay in control?

The problem is how that evidence gets assembled. The data is scattered across batch records, deviation logs, stability studies, LIMS, and certificates of analysis — and no one owns the integration between these systems, because the report only has to land once a year. So every year a team of quality engineers reassembles it from scratch — four to eight weeks per product, fifty to a hundred products per manufacturer — work that’s careful, necessary, and almost entirely mechanical.

Here’s the part that bothers us. The same scattered data contains the signals of failures about to happen. Stability trends that are drifting out of spec three months early. Deviations clustering on a single critical process parameter. Impurity outliers that only become obvious when you look across products, not within them. Nobody sees these signals, because the work of pulling the data together is manual and annual instead of continuous.

The APQR is the pharma industry’s most visible symptom of a quality data problem it doesn’t know it has.

Independent estimates of the cost of poor qualityin pharma — scrap, rework, investigations, recalls, lost batches — run from single digits at well-run manufacturers to 25–40% of revenue at those with weak quality systems (per the NSF and the American Society for Quality). That’s not an AI thesis; it’s an operations thesis. We think the reason those numbers have stayed so stubborn for so long is that the data required to prevent them has been locked inside a process nobody designed — it just accumulated.

What we’re building

Veriqa is quality intelligence for pharmaceutical manufacturing. Quality intelligenceis the layer that sits on top of your QMS and LIMS and turns the quality data they already hold into continuous, cross-product signal: where a QMS records events (deviations, CAPAs, change control) and a LIMS stores results, quality intelligence connects them and tells you what they mean. In plain English: we ingest the scattered quality data — batch records, deviations, stability studies, certificates of analysis — directly from source systems and documents, and turn it into a living, cross-product, trend-aware dataset that a QA team can actually query.

The APQR is one output of that system. An agent drafts it — with QA review at every decision boundary, not at the end. Regulatory-grade, fully auditable, and, critically, not a black box. Every citation in the draft traces back to the batch, deviation, or stability study it came from. If our agent gets something wrong, a reviewer sees it and overrides it in seconds. If it gets it right, a reviewer confirms it and moves on. This is the part pharma readers should care most about: nothing ships to a regulator without a human signing for it.

The bigger output is the one you don’t see on paper. Because the data is continuously assembled, we can surface signals before failures — stability drift, deviation clustering, CPP outliers across products— in time to do something about them. The report becomes a byproduct of a live system, not the reason it exists.

The system is an agent pipeline with human review gates built in — we’ll walk through how it actually works in the next post. For the regulatory side — what EU GMP Annex 22 changes for AI in pharma quality — start with our companion post.

Where we are, and what’s next

Today, the agent pipeline is drafting APQRs with human-in-the-loop review in production with our first design partner. One of our earliest design partners is Hema Pharma— where Margish’s family has run operations for over three decades. That proximity is how we understand APQRs from the inside: we get feedback on every quirk of deviation coding, every edge case in stability extrapolation, every practical judgment call a seasoned QA team has already solved. Alongside Hema, we’ve signed eight letters of intent with mid-size manufacturers.

Near-term, we’re expanding ingestion from spreadsheets into raw source documents — certificates of analysis, batch production records — so customers don’t have to prep their data before they see value. The goal: a new manufacturer connects their systems on a Monday and sees their first cross-product quality signals by Friday.

The real product isn’t the report. It’s a system that turns every manufacturer’s quality data into a continuous signal.

If this is you

If your team is burning weeks on APQRs — or if you suspect your quality data is telling you something you can’t hear — we’d like to talk to you. We work with a small number of design partners deliberately, because the early conversations shape everything.