KENX AI in GxP East 2026

2026-09-15-1702-gorawali-ai-validation-governance-manufacturing-lecture

48 min read

Gaurav Walia on AI validation and governance in life sciences manufacturing

Recorded 2026-09-15 at 1702 via Plaud. 49:53.

Summary

Closing lecture of the afternoon track at the GxP and AI conference. The presenter, Gaurav Walia, opens with his background: 30 years in industry, scientist by training, former director of R&D and later of quality at several companies including Pfizer, over 100 FDA inspections and three major consent-decree projects, trainer of 85 FDA inspectors over the last three years, now VP of CSV, CSA, and AI. His frame is a four-step innovation arc (creating, connecting, continuous learning, and development-and-deployment through MVPs and pilots) set against a data explosion, with 90 percent of the world’s data generated in the last two years and clinical trial volume quadrupled since 2008, while 50 percent of life science remains paper-based. He walks through quantified use cases: AstraZeneca improved visual-inspection defect detection by 70 percent; Pfizer used two AWS tools for predictive maintenance, cutting unplanned downtime 20 percent, and a separate multi-site program achieved a 55 percent infrastructure reduction with 16,000 cumulative hours saved; Merck cut vaccine-manufacturing process variability 50 percent; J&J analyzes production documentation, even scanned paper records through OCR, in near-real-time to catch procedural deviations, illustrated by a plant whose batch release stretched from 12 to over 32 days on documentation practice failures; Novartis used agentic AI for a 30 percent reduction in compliance documentation; Walmart’s pilot connecting ERP, point-of-sale, and weather data saved 86 million dollars; J.P. Morgan Chase runs mature AI fraud detection; Delta Airlines does predictive maintenance on flight and weather logs; and Alto Neuroscience used AI-defined brain biomarkers to predict treatment response in MDD and PTSD and progress a Phase IIa program. The unpublished core is an AI-driven CAPA demonstration: AI trained on ten years of historical quality data compares human-assigned root causes, which default overwhelmingly to training and human error, against its own top-three hypotheses with confidence scores, and the human investigator retains the final decision.

On governance he reports that the government is already in a beta cycle of an AI inspection tool for at least one industry segment, and relays a secondhand account of AI surfacing cross-site commonalities behind a microscope maker’s recent warning letters. He warns of dual-use risk, citing a published attempt to use Claude to enhance an African virus, and notes AI cybersecurity incidents roughly triple in the six months before elections. His regulatory argument: GAMP’s recent second edition calls for tailored AI validation beyond the traditional waterfall model; checkbox-style vendor audits are inadequate, and larger US companies are pushing enhanced quality agreements, SLAs, and defined “AI inspection” criteria for technology vendors; CSA let one CDMO implement a SaaS EBR/MES in three months; and the AI lifecycle differs structurally, with training models, data preparation, and orchestration software promoting tested models into production before formal testing. FDA expects explainability of what happens to the data; no black boxes. He stages an interview with a Moderna data scientist from a pure research background who has never faced regulations, using it to argue that discovery, GMP, quality, regulatory, and programming teams speak different languages and need deliberate bridging. On accuracy he flips the question: humans are far less accurate than assumed, and 25 years ago robotic-system validation was accepted as equivalency to human performance.

Plaud AI notes

App-generated, verbatim from Plaud’s Summary tab.

Date Time: 2026-09-15 17:02:51 Location: [Insert Location] Instructor: [Insert Name]

Summary

This series of lectures, presented by Gaurav Walia, a speaker with 30 years of experience in R&D, quality, and IT, provides a comprehensive overview of the application, governance, and validation of Artificial Intelligence (AI) in life sciences and manufacturing. The lectures address the exponential growth of data and how AI can enhance quality decisions, accelerate drug discovery, and optimize manufacturing processes. Key topics include real-time quality monitoring, documentation analytics, regulatory compliance, predictive maintenance, and AI-enabled clinical development. The speaker emphasizes the importance of bridging communication gaps between discovery, GMP, quality, and technical teams, and highlights the need for tailored validation approaches for AI systems, ensuring data integrity and explainability to meet FDA expectations. Several real-world use cases are presented, including J&J’s near-real-time detection of production documentation errors; Novartis’s 30% reduction in compliance documentation using agentic AI; and predictive maintenance programs at Pfizer and other firms that led to significant reductions in downtime and infrastructure costs. The lectures also cover AI’s role in drug discovery, citing a virtual AI company’s success in selecting drug candidates and Alto Neuroscience’s use of AI-defined biomarkers to achieve positive Phase IIa results in MDD/PTSD trials. Furthermore, the discussion extends to AI-driven CAPA processes, leveraging historical data to improve root cause analysis while maintaining human oversight. The speaker also addresses the dual-use risks of AI, the need for robust ethical and security frameworks, and the evolving regulatory landscape, including GAMP guidance, CSA, and forthcoming changes like Annex 22, while underscoring the increasing AI-driven cybersecurity risks as of September 15, 2026.

Knowledge Points

1. Introduction and Speaker’s Background

  • Speaker’s Experience
    • The speaker is Gaurav Walia, who has 30 years of industry experience, including roles as a Director of R&D and Quality at companies like Pfizer.
    • He was involved in developing the first moderate to late-stage Alzheimer’s medication and has been through over 100 FDA inspections and three major consent decrees.
    • In the last three years, he has trained 85 FDA inspectors on identifying industry gaps.
    • Currently, he is the VP of CSV, CSA, and AI, heading two large North American offices.

2. The Role, Innovation, and Evolution of AI

  • Four Basic Steps of AI Innovation
    • Creating: Brainstorming applicability to solve problems and gain insights from the massive amount of data generated recently (90% in the last two years).
    • Connecting: Socializing ideas, building an innovative ecosystem, and engaging with tech vendors, who often lack the real-world data that companies possess.
    • Continuous Learning: The evolution of AI models, which may lead to future applications like autonomous approvals, requiring regulatory changes.
    • Development and Deployment: Starting with a Minimal Viable Product (MVP) or pilot to test use cases and gain confidence for larger deployments to maximize ROI.
  • Evolution of AI
    • The term “AI” has existed for over 50 years, gaining public awareness with events like IBM’s Watson on Jeopardy.
    • A key breakthrough around 2017 showed AI could diagnose certain cancers from scans up to six months earlier than approved diagnostic tests.
  • AI vs. Human Accuracy
    • The speaker challenges questions about AI’s accuracy by pointing out that human accuracy is often much lower than perceived, though data to prove this is scarce.
    • 25 years ago, the validation standard for robotic systems was equivalency to human performance.

3. Validation, Compliance, and Data Integrity in Regulated AI

  • Bridging Organizational Silos
    • Significant communication gaps exist between discovery, GMP, quality, regulatory, and programming teams due to different terminologies and concepts.
    • Cross-functional collaboration and a shared language are essential for AI projects to meet regulatory expectations for validation and data integrity.
  • Tailored Validation and Regulatory Frameworks
    • GAMP guidance suggests a tailored validation approach for AI, moving beyond traditional waterfall models. Insufficient vendor audits (e.g., simple checkboxes) are inadequate for complex AI technologies.
    • AI validation must account for training models, algorithms, and data transformations before standard testing phases.
    • Explainability is crucial; FDA requires traceability of data transformations to avoid “black box” issues during inspections.
  • CSA-driven Acceleration and Vendor Oversight
    • Computer Software Assurance (CSA) enables a risk-based approach that accelerates the adoption of SaaS and cloud systems like MES/EBR.
    • Enhanced quality agreements, SLAs, and defined “AI inspection” criteria are critical when working with AI technology vendors.
  • Distinct AI Software Lifecycle
    • The AI validation lifecycle includes pre-testing steps like model training and data preparation. Orchestration software can be used to manage the promotion of tested training models into production for formal validation.

4. AI Applications and Use Cases Across Industries

  • Data Explosion and Market Trends
    • 90% of the world’s data was generated in the last two years, and clinical trial volume has quadrupled since 2008.
    • Despite this, 50% of the life science industry still uses paper-based processes, and supply chains are increasingly complex and vulnerable.
  • General AI Use Cases
    • Visual Inspection: AI with vision systems improves defect detection, replacing tedious manual inspections where human performance degrades. AstraZeneca improved detection rates by 70%.
    • Complaint Handling: AI can triage queries and complaints, though there is a risk of critical misclassification.
    • Document Management: AI assists with standardizing company-wide procedures.
  • Pharmaceutical and Manufacturing Use Cases
    • Real-time Monitoring & QC (J&J): AI analyzes production documentation (even from scanned paper records) for procedural deviations in near-real-time, reducing batch release times and risk. Systems can also monitor physical specs like dryness percentage for dynamic process adjustments.
    • Predictive Maintenance (Pfizer): Used AWS tools to predict equipment anomalies before failure, reducing unplanned downtime by 20%. A separate multi-site program achieved a 55% infrastructure reduction and saved 16,000 cumulative hours.
    • Process Optimization (Merck): AI analysis of manufacturing data for vaccine production resulted in a 50% reduction in process variability.
    • AI-Driven CAPA: AI analyzes historical quality data (e.g., 10 years) to identify root causes more accurately than humans, who often default to “training” or “human error.” The system provides top root cause hypotheses with confidence scores, but humans retain final decision authority.
  • Clinical Development (Alto Neuroscience)
    • An AI-enabled platform using EEG, behavioral data, and genetics predicts treatment response for MDD and PTSD.
    • In a Phase IIa study for ALTO-100, biomarker-defined patients showed greater depression severity reduction, enabling successful progression of the program.
  • Cross-Industry Examples
    • Financial Services (J.P. Morgan Chase): Mature use of AI for real-time fraud and anomaly detection.
    • Automotive/Aviation (Delta Airlines): Predictive maintenance using flight and weather logs to anticipate component failures and schedule maintenance, minimizing downtime.
    • Retail (Walmart): Used AI to connect ERP, sales, and weather data to predict supply shortages, saving $86 million in an initial pilot.

5. Dual-Use Risks, Ethics, and Governance

  • Dual-Use and Security Concerns
    • The potential for AI to be weaponized (e.g., designing an enhanced virus) highlights the need for robust controls, third-party verification, and potentially slowing deployment to ensure safety.
    • Cybersecurity risks involving AI are noted to triple approximately six months before elections.
  • Ethical AI, Governance, and Organizational Readiness
    • Implementation requires ethical AI principles, comprehensive guardrails, and strong security/legal foundations.
    • Organizations must stay current with regulatory changes (e.g., Annex 22 draft) and build readiness through talent, processes, and digital validation efforts.
    • Collaboration in AI subcommittees provides insights not yet public, informing prudent adoption.

6. Workforce and Practical Considerations

  • Consequences of Poor Documentation & Staffing
    • An example highlighted how cost-cutting and hiring non-native English speakers led to severe Good Documentation Practice (GDP) issues, increasing batch release times from 12 to over 32 days.
  • Skills and Regulatory Exposure
    • There is significant variance in regulatory exposure among AI developers. A developer at Moderna with a research focus had no direct regulatory experience, underscoring the need for compliance education as teams move from research to regulated development.
  • Global Operations and Data Aggregation
    • Global companies face challenges in aggregating data from diverse sources (devices, wearables, apps) and must harmonize validation and data integrity frameworks across jurisdictions.

Questions

  • [Insert Question/Confusion]

Assignments

  • 1. Establish a cross-functional glossary to bridge communication gaps between discovery, GMP, quality, regulatory, and engineering teams for AI projects.
  • 2. Define and implement a tailored AI validation framework aligned with GAMP guidance, covering training data, model versioning, algorithm explainability, and traceable data transformations.
  • 3. Develop enhanced quality agreements and SLAs with AI technology vendors, specifying criteria for AI inspections and security expectations.
  • 4. Adopt a CSA risk-based approach to accelerate the deployment of SaaS solutions like EBR/MES, leveraging vendor documentation and quality audits.
  • 5. Pilot an AI-driven CAPA initiative using at least 10 years of historical quality data to improve root cause analysis and measure improvements in investigation time and effectiveness.
  • 6. For clinical programs, assess the feasibility of using AI-defined biomarker platforms to stratify patients and enhance efficacy signals, drawing on Alto Neuroscience’s methodology.
  • 7. Develop and formalize an ethical AI and security framework with guardrails, third-party verification, and legal oversight, including enhanced cybersecurity measures around election periods.
  • 8. Track regulatory developments (e.g., Annex 22 draft) and participate in digital validation/AI governance committees to align organizational readiness with industry direction.
  • 9. Evaluate current production documentation workflows; consider implementing OCR + AI triage to detect gaps and root causes in near-real-time to reduce batch release delays.
  • 10. For global operations, design a harmonized data integrity and validation standard to accommodate multi-site, multi-country regulatory requirements.

Transcript

[00:00] Speaker 1: Probably a long day for everybody. I’m often the last speaker on day one, and I hold you for your drinks, so I’ll try to keep it as brief as possible. There’ll be some things that I will talk about that I can’t share publicly, but I can talk about the concept of investigation, CAPA, root cause analysis with an AI, and how it operates. Some of this is not quite public yet, so some studies are still going on. But I can talk to it anonymously. Today we’re going to look at a little bit of sharing about different use cases of AI across different industries, not just life science. We’ll look at the data explosion as well as some of the trends in the market. You know how can we really enable the use of AI to help with quality decisions, quality metrics, and some of these may highlight some, you know, one slide use case of of AI quality metrics in manufacturing or utilization of vision system or other areas. Okay. So quick background. I think I’ll give you a little bit of the extended version from this morning. I’m Gaurav Walia. I’ve been in the industry for last thirty years. I’m a scientist by training, former director of research and development for several companies, including Pfizer, and and so I get bored easily. It’s always been my nature. It means that I have to have my hands in many things. So as a scientist, I very quickly kind of got bored of the science, and it was just another product with the same technology, similar methods, and. It was exciting when we developed 25 years ago the first moderate to late stage Alzheimer’s medication, as well as some antidepressant drugs that were novel. But 30 years ago was virtually small dosage, small molecule. And today we have a better avenue of medications and delivery. I’m also former director of quality. I’ve been through over 100 FDA inspections in my career, as well as three major consent decree projects for large pharma, mostly in the space of data and CSV across some companies where their sites were shut down for five years and couldn’t sell any product. You probably heard of metal contamination in baby aspirin some years ago. One of the top pharma companies in the world, just as an example. And then from there, in the last three years, I’ve trained 85 FDA inspectors. With the second in charge of our company for the quest, and it was really a fusion of experienced inspectors with the newer inspectors, right? And really coming out of COVID, it was trying to look at what are the common common gaps in the industry, whether in the lab, the manufacturing, enterprise system, and what are the gaps that inspectors should see when he walks up to a A configuration screen, a control panel, a scan system to be able to deduce that. So, today I’m the VP of I’ve been a regional director of IT as well, and today I’m the VP of CSV, CSA, and AI, and I head two of Group’s largest nine North American offices. So, if we get into it, you know, AI is about understanding some steps, four basic steps, which as you’ll see shortly, and how do we apply the innovation? Not just equality, but to our everyday life. How many people are here? Raise your hand. Are quality people? Raise your hand. Good amount. What other categories do we have? Anybody in the lab? Not a lab conference, but anybody in manufacturing specifically? Did I miss anybody? Anybody in a special role? Product management. So what are you managing specifically in the product? AI software for the manufacturing. Do you mind me asking what kind of product? Pharmaceutical medical device.

[04:26] Speaker 2: No. Originally we worked with manufacturing companies like SanDisk and Drucker, doing operational manufacturing transfer, primarily through video training. Now we’re getting more into the physical AI space with PLM. Good.

[04:45] Speaker 1: Thank you for sure. As you heard, an example is how to leverage AI. Some of it’s about acceleration. Some of it is about acceleration in drug discovering. If I tell you a quick story, in the last two weeks, a new article has come out, and there was a virtual AI company that created an AI environment with thousands of AI agents. They started with creating the physical structure of an organization, from the quality to the scientists to the chief scientific officer of the company. It even had a virtual environment that looked like you see in the room today. And then it created its agents that were specifically purpose for drug candidate screening as well as selection. It picked the candidates that thought were the best from the data it had, as well as from its postulation what would what would become commercially viable, and ultimately in phase development, of course, not full commercial release. But throughout the phase process, it evaluated and selected these candidates, and in the end, I believe Merck did a real life. Using humans, using their normal process without AI, to deduce the same, and the results were identical. So, though it’s quite interesting in the proliferation of innovation, in order to not only do the research and develop develop quicker pathways for becoming the first in the market or the first best in the market, and by using innovation tools. So, if I go back to the four steps, it’s about creating. Right. I mean, even behind creating is the brainstorming of where that may be applicable. Right. The creation involves looking at solving a problem, meeting unmet needs, being able to look at insights of the data you have. Today, there’s a statistics that just like somebody this morning said, like the internet doubles every six minutes. They say that the in the last two years. Ninety percent of the world’s data has been generated in the last two years alone. I don’t know where that statistic comes from, but well, it means that that there’s there’s large volumes of data that are not even comprehensible in a short time by human. Right. That means that we need to we need to solve problems to be able to analyze the number of clinical trials that have increased since 2008 has quadrupled. Right. That means the sheer volume of data. Before clinical trials, these manual processes, right? 50% of the world today, believe it or not, in life science is still predominantly on paper. So there’s many steps for companies to evolve to AI. It means that you have to first digitize, and you have to adopt technology, and then you can start taking steps towards processes and data consumption besides typical processes. The other is to connect, to socialize, and build an innovative ecosystem. Right? We talk about agentic AI. Probably two years from now, we won’t be talking about it. Right? So the evolution is fluid at this time. It means that the pipeline has to think about where the road may go. It means that it’s it’s important to engage with technology vendors, big and small, to see where the innovation might be going. Because it’s somewhat of a guess, right? Do you pick the right plan or the right pathway? Oftentimes, the the problem with innovation of tech vendors is they don’t have data. They don’t have real data, right? Many of you have the real data, and that’s a challenge because it’s not typical validation where you just test inputs that should produce known output and that’s it. This is a different game. That means there’s a whole other set of learning with the data itself, depending on the application. That’s going to tell you if it’s right or not out of the box, your box, different out of the box. And then, of course, as you’ve heard today, continuous learning models. I think in the future there’ll be there will be continuous change control monitoring that will be utilized by AI agent. Even even in the future, there will be approvals of I think autonomous AI. It just means that some of the regulation will have to change as technology advances, right? And we have confidence in the output. So if we look at development, we want many companies talk about minimal viable products. What is the first step I want to do? What pilot do I want to create? What are the business use cases today? I would say. Manufacturing is probably one third of the of the pilot’s quality is the other third, and discovery is the other third. There’s, I know that’s that’s close to 100%. There is a niche also specifically in in AI enabled medical devices. So, but that follows a slightly different track based on both regulation, new ISO standard, as well as the newer FDA draft guidances from the last year and a half. So ultimately, the deployment becomes important not only for the POC or pilot, but in terms of we’re not here to do a small study. We’re here to produce some confidence and produce some valuable results to either continue or not, and carry that message forward. Ultimately, companies eventually are looking for bigger deployments, where you can maximize the volume of utilization to maximize your full ROI, assuming we’re talking about better quality, savings, efficiencies in particular systems. In drug development, we look at being able to put forward the best novel candidates, best in class, and to beat some of our own competitors. If we look at some of the timeline in the middle there, we see that AI has been around. The terminology has been around for 50 plus years, right? It wasn’t really until we started getting some social aspects of it with IBM’s Watson beating Jeopardy, and and the first time I knew about AI was probably five years after 2012. I read some of those studies that were able to diagnose cancer using scans and different imaging data, and ultimately, what they found was the AI of through cancer imaging recognition and machine learning and some AI was able to detect certain forms of cancer six months earlier than any known approved diagnostic or. Or tests in the world, and so for some people it saves their life. So ultimately, what we look for in the development depends where you’re looking at. In some places, if I ask, if I tell you 25 years ago. The conversation we have in research was asking different regulators and industry folks, “How do I validate a robotic system?” And at that time, what we said was, “You need to validate it by showing equivalency to the human.” I mean, literally, we’re humans. We’re taking syringes and sampling in the lab, putting it another machine to analyze. It’s quality, it’s purity, it’s impurity, and so forth. So it’s interesting with AI. Often we hear conversations about 95% accuracy and 100% accuracy. Is the human 100% accurate, or even 95? If you even have human data, which most companies do not, I guarantee you, if you don’t already know, the human is much less accurate than you think. It’s an interesting paradigm of what that number should be, and what’s practical, and what’s not. And it depends on your system, the risk, and the how. I do. So this is just some highlights of the data explosion. On the top left, you’ll see about what I was talking about earlier in terms of the amount of data that’s generated. Of course, today we have truly global supply chains that are complex, and oftentimes in the last year we had some clients. I’ll keep them anonymous, but it’s public data. But where a dam broke or a natural disaster occurred, and their facilities were down for six months, and there was significant drug supply shortages with those organizations. Where we help those organizations get back online, but in the meantime, what do you do when some of those are critical or they dominate the market in that particular product? I worked on some consent decrees where. The company had 60% of the world’s market, so it becomes difficult. Then it becomes a challenge because you can’t just stop supplying the world for five years of products. So it becomes a negotiation with the government, right? And so, of course, this this boils down to regulatory pressure. Nobody wants to have a consent decree or have a major product issue, and depending on who is affected or who the main users of those products. It’s it’s brand imaging, right? Brand imaging has a direct impact to the, especially if you’re a public company, right? And so, and the confidence in the overall brand, right? No matter how many business lines or products you have, they see the one name. So, the demand for demand for better quality becomes even more critical in the pipeline of of evolution, and so, so I think as you’ll see next in some examples. It’s about early detection and potential issues, and being able to analyze that large volume of data to make better decisions and even more preventative. This is kind of ultra preventative, more or less. In these examples, I show you one case study that’s outside of the life science industry. One key point: if you look at it, just read this. Number one, 2019. Walmart basically wanted to reduce the amount of food waste in Europe, and what they did is they connected their ERP systems to their point of sale systems to weather conditions that would cause either supply shortages, like there’s a cold streak in Florida that all of a sudden you have a shortage of oranges, then suddenly the price goes up or the availability of that product goes down, and so that has an impact. In its first use case, they saved eighty-six million dollars in the first pilot of AI within Walmart. So these are this is a thirty-thousand-foot view. You’ll see some examples of this when it comes to AI use cases over the last few years. But if we see overall, we see we’ll see a lot of examples of AI used in with visual inspection systems for product defects, as well as. Some companies, believe it or not, still have humans looking at even right where I’m in the city where I’m from. I’ve been to some plants where their process is the human inspects every bag, right? And so, one, it’s a tedious task. Two, there’s usually operational performance that I’ve found historically after four hours of doing the same task repetitively. And so you tend to have more investigations, defects, and you’re only lucky if you catch it before it goes on the bill, because then you have its own damaging, effective issues. So even complaint handling, there’s been a number of organizations that look at not only the triaging of a query versus a complaint, but also what’s the level of action that can be escalated after initial AI looks at it. By a human. But that becomes tricky because if it fails and it made a critically wrong decision, then what’s the impact of that result? Is it a death of the patient? Is it missed something somehow? I think none talk about omission earlier today on the panel, but ultimately, some people are going with. Some of the administrative, I call examples of document management, reviewing. I think some of some of the folks met with privately yesterday talked about the quality excellence of standardizing all of the procedures to one set of procedures for the company, no matter which side. And I think it was a precipice for them to standardize the procedure, then to. Further implemented, so there was one way of doing business, right? Which most companies have global policies and procedures, and they don’t line well sometimes, depending on the organization to the site, right? It’s I’ve been at companies that, unfortunately, if you were to count and hum it on a couple hands, how many people are actually fully trained in those global policies and procedures at the local site? It’s usually quite deficient at some companies. So, so if you’re not, you don’t even know it exists. How can you align to it? So the the outputs that we’re looking for in some of this is looking at obviously quality improvement, cost reduction, efficiency gains in terms of quicker batch releases with less problem problems, as well as what I said before, overall customer satisfaction. Right? You can have different ROIs depending on the use cases, and you can also improve. And I’ll give you some insights. Regarding the agency that I’ve learned in the last month or so, based on some different world-leading companies and what they want to try to do, and then I, we have some people on our team that I interview that are ex-FDA, and I came to find out some things I didn’t know. I’ll talk about that in a minute. So these are a couple high-level examples, right? This, this is as we drill a little bit further down. If we look at Astrogenica, recently they performed an AI-based visual inspection that combines with their visual inspection system to improve defect detection in their products. In this case, they were comparing it to the man, their current process, which was manual inspection of their products, and they were able to reduce the amount of defects by seventy percent in the first. I don’t know if it’s a pilot or PLC, but, So it helped the overall process as well as the the yields, and and when they had issues, at least it was further evaluated close to real time, right? The other is it’s a famous example of predictive maintenance. I may have another slide on this. I’m not sure. So Pfizer used two different two different AWS tools that are commercialized, and the idea is to be able to tell well in advance when something is behaving out of not even. Failing, but out of out of character, out of trend. In order to kind of see what’s an anomaly, right? And in order to be able to properly plan that, maybe maybe we have to do a maintenance sooner. Maybe maybe there’s better planning behind the team and the production of that particular product or line. Merk in the next example used process automation. In looking at manufacturing data to optimize process parameters. So, as an R&D scientist and director, 25 plus years ago, I remember at that time we were implementing PAT or process analytical technologies in manufacturing, from lab first and then to manufacturing. We said if we can add a sensor within each step of the manufacturing process, we can almost have an online chef. That can that can look at what the specs and operational conditions are, and maybe maybe the the spec is that it has to be five, it has to have a certain percentage of dryness, and at that it can monitor real time what that dryness is, and it can shorten the time based on the specification of that that physical characteristic, right? And ultimately, it also can furthermore down the line can online QC. Some of the some of the specs against the physical properties of that product, right? So, the idea is to be able to monitor in better real time. In the next example, if you look at J and J, J and J was able to look at quality attributes in production. So, what does it mean? It means that very simply, they’re looking at documentation, right? They’re looking at what mistakes are being made when the procedures not being followed when the batch. When the batch method is not being followed, so they can catch almost in real time. I can tell you from real life experience that I’ve experienced even some of the largest companies in the world, believe it or not, that started cutting their costs in hiring people in the plant that were not English as the first language, and they just speak the truth. In the end, they were having a multitude of GDP issues where it says. Oh, it’s missing this product or material, or it’s written so vaguely that the quality people can’t release the batch. And so suddenly, this this company had fifty, sixty sites in the world. This site was that was local to where I’m at was the second worst in the world. The amount of time it took to release a batch was when normal release periods were twelve days. It took it took at least thirty two days. Right, and that means it’s at it’s at a high visibility of the C-suite. It means that the product life is decreased. It means that there’s not confidence for what’s a really simple mistakes. Right, I mean, you can easily put an OCR AI on top of even if you have paper on top of that. That can identify the gaps in the documentation real time or close to real time because you have to scan it if it’s paper today. And and be able to at least triage what’s going on and what’s the root causes. That way, at least the risk becomes reduced. You nullify the problems quicker. When you have a batch record hanging on a wall for for weeks and it doesn’t get resolved properly, it’s a problem. And it wasn’t just about the GDP. It was about the interpretation of what it was trying to say or not trying to say. And that interpretation was unclear. That the people couldn’t release the product. So, if we look at regulatory compliance, Novartis created a 30% reduction in compliance documentation. So today, many companies have agentic various agents that will look at what they will have separate agents to look at GDPR, COIL, data integrity, cross-reference the company policies and procedures, regulations, as well as the The hard conditions that are needed to have good documentation to ensure regulatory compliance. Right. Similarly, if I share with you, I’ll jump a little bit ahead. But so maybe not officially in terms of public, but I can tell you that the government already has been creating and utilizing AI for during inspections to analyze data. Okay. I met with somebody recently that trains FDA inspectors as ex-FDA, and we had a client that that wanted to use an ex-FDA inspector with an AI, so they get the full picture of what the gaps are for a medical device for a large one. And the idea was that they wanted to see how mature. Not only their he had taken a VP had taken over a new part of the organization. You want to see how mature it was, but They also were encountering severe gaps in acquisitions where these companies were being inspected, and they were they were failing pretty miserably in the in the new ISO 1345 inspection standard. So the idea conceptually here was to create or not create, but utilize an AI with the human to to really see the full picture, because maybe the AI misses some things or it’s not perfect yet. Maybe the human. In this case, I wouldn’t think so, but the human can also make mistakes. And overall, what he told me is that the government is already in beta, in beta cycle of for that specific set of the industry for an AI for an inspection tool, right? That that inspectors could utilize to analyze your data. I’m starting to see. I saw some. I won’t name the company, but I saw some warning letters that came out this morning, and I happen to know about it because. A guy that used to work for me many years ago. His wife works in that company, and told me that that company that makes microscopes. And traditionally, a lot of microscope companies have a lot of compliance issues. They don’t have good software. The microscopy technology is quite good, but it traditionally gets cited often. And so, this company has several warning letters across multiple regions of the world. And I don’t know officially, but the rumor is that there was AI not only helping to find some of the issues with inspectors, but showing commonalities across multiple sites of the company in different unique regions of the world. So I’ve worked for some firms where the fear becomes what we call systemic regional issues. It’s not a bigger organizational problem in those companies. It’s not a one-off. It’s not a department. It’s not a site. It’s bigger than that. This just added more for take-home. You get copies of these slides afterwards. This is the one that was used for predictive maintenance in the first year across multiple sites. They had a 55% reduction in infrastructure. And cumulatively across multiple sites, say 16,000 hours. Downtime becomes important in manufacturing, whether planned or not. This is another case study. I just wanted to share it with you, and then maybe in a few minutes we’ll get into some of the quality metrics upon Kappa’s investigations. Again, I don’t have slides for that because I can’t publicly share it. I can talk about how the tech works today. This is all to neuroscience. They were clinically developing the product in the clinicals, and they were for PTSD and severe depression. They used AI to deduce which patients might have the best results based on certain brain biomarkers. If you see, ultimately their end result was. They were able to move it on into other phases of development due to greater reduction based on those biomarker indications for that particular drug. Unfortunately, at the same time, I share with you this is about AI and drug discovery and advancement. In the last week, it was published that somebody tried to use Claude to develop an African virus to make it super potent and even more deadly. One thing we haven’t talked about too much here, but I tell you because I talked about a little bit on the panel, but it becomes important because you know some of the leaders of AI companies are saying it has to have third party verification, as Dory said. It needs better controls. It needs we need to slow down. I mean, there’s two sides of the reality. The one side is that there’s the basic reality of it’s a race. AI is a global race. AI also, what we don’t hear is what the governments are doing with it, right? And so, in a way, it can be weaponized. Just like if you can discover a drug, you can make something that’s lethal, that’s a virus in the community for bad actors, right? We usually think of bad actors in cyber, but As you heard earlier, it is being utilized for that as well. So, whether we have defense systems in the military that can counteract, this this what goes on in the world. And so, the framework of controls in the life science industry are important to have. And you know, you just you can’t create a bullet train overnight, right? You need to you need to have the infrastructure in place. And has anybody ever ridden a bullet train? Raise your hand. Okay. You have to get your bearings if you’ve never been on one, right? You just don’t stand up and walk through the train, do you? Right? Because it’s moving at a speed that’s not what we’re used to, and and maybe a bad analogy, but I consider AI to be that as well. So if we look at some of the framework, right? The left side means some of our clients start with ethical AI. They start with framework. They start with guardrails. Including all the different types of data and how it should be controlled, right? Some start with more the security and the legal because there’s public companies have a lot have a lot of risk. I don’t know if you know this. Six months before an election, the amount of AI cybersecurity incidents increase like triplefold, and if you’re public and you’re a top 10 or 20. Company or in the news, they’re even more risk, right? And so, so the other side is the regulatory compliance. Tomorrow, I’ll give a talk about the current, current acts, guidelines, guidances, and what the future is, what’s coming in the future, right? And I think you’ve heard some of that today. Many of us sit on the same digital validation subcommittees and AI subcommittees. So we get to hear some things that don’t get publicized, terms of industry and government’s meeting and working through the process, right? Such as you know the the problem with the current draft version of Annex 22 and its limitations. So the other, as you’ve heard us talk about today, is organizational readiness, right? Anybody a programmer in this room or actually a developer that? Writes code for AI. You’re the only one. Good. And let me ask you, what’s your background, and what kind of company do you work for?

[34:25] Speaker 3: I work for Moderna.

[34:26] Speaker 1: Okay, I know it.

[34:27] Speaker 3: I did biological engineering for grad school. I taught myself programming.

[34:35] Speaker 1: Is that your first job? What did you do first when you entered the industry?

[34:40] Speaker 3: I was at Novartis. What were you doing? Data science. So Moderna data science as well.

[34:46] Speaker 1: I forgot to. How was your journey in terms of compliance and regulation in the framework of developing tech or with data?

[34:57] Speaker 3: I’m a complete newbie. I’ve been in research all this while, preclinical, so no regulations.

[35:05] Speaker 1: Yeah.

[35:05] Speaker 3: Then I didn’t have to.

[35:08] Speaker 1: So why did you have to do that? Because you know, my my company used to do that too, and I said, hey, you know, when you have a polymorph crystal and then you want to file a patent so you can extend the life of the product, you know, if you have no GMP, no documentation, you got to go back two years and do it over again. I hope you remember, right? And that’s the reality, right? Like it’s a different brain, and and oftentimes. Whether we’re scientists or developers or quality people or regulatory people, you know, almost sometimes with AI especially, there is a. I remember I was going to take a discovery job, and discovery people can’t speak to GMP people. They speak different languages, almost like you’re speaking a foreign language. The terminology is different. The concepts are different. There’s no bridge to connect between the groups, and sometimes, you see that too. Right, a programmer may not be able to speak compliance. Compliance guy can’t speak to the programmer. They’re speaking two different languages, but we work together as a team to create that harmony. So, I think all these become important. Validation itself, you have you have regulation that’s been around for twenty years. You have GAMP AI and different frameworks and recommendations, and even your own your own. It means there’s no formal framework. There’s ideas, right? The government then eventually will come up with its own to inspect us against, and and then of course the future. The future is a big question because, as you know, regulation takes a long time. The tech is moving too fast. It will it will stabilize at some point, but and who knows when regulation will come? And and there’s also the standardization between governments because many of us just a quick show of hands. How many of you work in a. Company that has multiple sites in different parts of the world. Okay, so good part of the audience. So, so the the issue becomes is how do you manage something with all this data, with all the different datas coming from devices, from apps, from wearable technology. If you have an AI-enabled medical device, or we know all the different products that might be in the market. You know, we start with Machine learning in this process. Most companies started with this, right? Even before that is the example of PAT. PAT was no different than machine learning. You you literally train it, the calibration curves, what’s good, what’s bad, and it understands it, right? And that’s it’s really what machine learning is today, more or less. It’s the deep learning and and the new technologies, neuronal networks, and ultimately. You know what the most common thing is discussed today is data integrity within AI because it’s a different version of data. It’s a different version of integrity. The integrity can mean different things. Not just am I capturing all the metadata? Am I not lying? Am I not doing fraud? I mean, there’s there’s different levels, and because because depending on what you’re doing, people can’t. Explain what happens to the data. Remember, if you ask, they’re not here today, but they were supposed to come. But when FDA is here, they’ll tell you. You need to explain what happens to that data for any time. Can’t be went in, went inside a black box, and got came out. Can’t do it. You’ll have problems during an inspection if you can’t explain because there is, there is. Data. There’s calculations. There’s algorithms. There’s things that go on inside that changes the data, and that’s part of the data integrity process when you do AI. This is just more of a take home, but there’s a couple slides. On the left is from Gap, and the second edition that was released recently, and it means that ultimately this all needs a tailored validation approach, right? You can you can if you want. I see some companies that. Come to me and say I have an AI, and we’d like to we’d like to do traditional waterfall model. So you could do it. Sometimes they even say it’s a SaaS system. We want to do the waterfall model. So you have vendor audit. Yeah, I think so. Three pages long, half it’s not filled out. If you really look at it, it’s like they added a checkbox for technology vendor evaluation, and it’s it’s really a product and material vendor audit, not. As you get into AI, companies, some of the bigger companies in the U.S. are focused on the the enhanced quality agreement and service level agreement, as well as what does an AI inspection look like for a technology vendor? I mean, of course, that has many layers, right? Of security, if you’re going to use that concept, some don’t. Some keep everything within their walls for today, which I understand. Even if we talk about CSA, this not a CSA talk, but this an example of risk, right? It’s an example of of different ways of testing. I I always call not only because of the cloud and the and the leveraging of technology vendor documentation and your quality audit, by by leveraging what CSA has allows companies to go. I’ve worked with some companies in the last couple of years that were on paper two years ago today. Today, in the last six months, they implemented CSA, implemented SaaS in an EBR MES system, and they did it in three months. Right? They didn’t use a. I wouldn’t say they used the top version of a of an MES system in the market, but something newer product, right? And let’s say a light version of of the EBR MES. But the idea was they wanted to move faster as a CDMO, and they wanted to have competitive advantage in the market. Why? Most of the most of clients are bigger companies. Most bigger companies, at least domestically, are not on paper. It means that they’re familiar with certain systems, and when you can communicate that data and the results, whether it’s an audit or it’s just transaction of data and sharing of of whatever the of the product data is, that becomes important, right? Using similar tools and methods, it becomes easier to check. It becomes more apples to apples. This is an example of of a software lifecycle ecosystem. It’s just an example; it doesn’t mean has to be one way. It means that you know you’ll see a couple slides, and it means that the the validation lifecycle is uniquely different, right? You may have some things that you do like you would do with Part Eleven in your typical CSD process, your CSA process, right? But the difference is going to be in your training model, right? And you have many steps. Before you’re actually running official tests, like you would in a typical CS3 model of an older system, right? And ultimately, some companies have orchestration software that will test the training models, but then also go into production of a training model in order to actually do the testing. So, if we look at a couple examples outside of life science, and I only share it just to. Show you what some of the companies and other industries are doing. J.P. Morgan Chase utilized AI for the last several years in fraud detection. Everybody has a bank here. People, the AI automatically triggers and shows there’s fraud or there’s some unusual activity, and you may get an automated call. You may get a human that verifies and goes through different steps to verify. Right? How many people in this room have had have had that scenario occur? Okay. Myself, right? So, so the financial sector and the banking sector have been utilizing AI for some time, and nobody wants to lose money. Every penny counts, and it’s ultra sensitive when it comes to financial regulations and the control of AI. But again, if you look at some of these things, it looks at risk analysis, document automation. It’s looking at how can I look at faster anomaly detection. So whether we’re talking about manufacturing or we’re talking about financial. And they’re also looking at pulse quality, right? You want to be able to make sure that that I’m not only showing what the AI can do, but what it can’t do. What are my limitations and boundaries? As you’ve heard today, if you look at this global automotive manufacturer, they wanted to look at predictive machine failures. I think I might have it here, but they’re also the the aviation industry does the same thing. They really have real-time data not only in the maintenance,s the mileage, but also when they’re seeing anomalies. In in some part on the plane, so they actually know before that plane lands that they need to do something. That’s why sometimes flights get delayed. I remember recently traveling, I heard the pilot whisper to the guy at the desk, “It’s going to be a half an hour delay. They have to do this maintenance.” And it was interesting because ultimately, when you don’t plan it, it has larger downtimes. When you don’t have a backup crew or backup plane, you can’t fix it properly. One time, was at airport, took two and a half hours to change a tire. Change a tire, right? That would be like a NASCAR pit crew that finds you know two two seconds in your living out, right? And it’s not that. So it’s it’s about efficiency overall and and knowing your process because you know the airline industry. It’s important that you have plus or minus a big number that you have on time flights, right? They get valued or not then. So. If we look at a couple other, this is the example of Delta Airlines. So Delta Airlines really wanted to avoid just ground disruptions and unexpected maintenance,s and reactivness in maintenance. So whether you’re fixing a plane or you’re fixing your production line and your medical device manufacturer, pharmaceutical plant, it’s it’s relevant, right? And if you look at this, it’s looking at flight logs, weather logs. I mean. Different, but similar to the Walmart example, they looked at weather conditions to take all that cumulative data to postulate what happens in the supply chain. What happens when with the cost? So, if you look at AI-driven capital, we have maybe just a few minutes left. So, I I’ll focus on the on the verbal part of this since I told you I could not share the slides I use. But so. A couple years ago, a company had come and they were interested in looking at historically. You know, today most organizations have a much higher turnover rate, which contributes to some human factor of the consistency of root cause analysis investigation, as well as many companies have an effectiveness check. But the reality is, in this study, what we did is we looked at historical data of ten years’ worth. They looked at the quality data in this line, and they compared the 10 years of historical data. What the AI does is the AI will analyze what the human saw was the root cause, what was the variability in the root cause, and did it really get the root cause correct or not? Some companies are better than others. Some their their investigations get closed out improperly. They have to hit timelines. There’s a huge percentage of training and human error. When the reality is that training is not the overwhelming issue, by intelligence laziness, and so it means that the AI would analyze the data. Then the AI on the right side will start based depending on how you configure it. Will give you however many options you you ask for. You say I want the top three. Root causes for this issue. Right, it’s going to look at your procedures. It’s going to look at your historical defects or issues. It’s going to look at your historical trends in this particular space. That probably in the beginning that align. All right, some you’ll find that they don’t get matched one hundred percent, and that’s where the human comes in. So, ultimately, what what this analysis will do it will tell you what it thinks the confidence. Of the human decision was, then it will deduce what it thinks based on the data and the history. What the AI thinks is that’s where the human then steps in. AI, the human will look at is it. The AI might say number one I think is 97% accurate. Number two is 85% accurate. Number three is is 72. And the human might say you know based on my experience. Based on the data, based on what I think it is, maybe it’s number two. It’s really not number one. So there’s some fine tuning that may go on, and what we yielded is better, better results in the overall process. Even if it’s just a tool, again, today the AI doesn’t make decision. Means that the human can at least digest 10 years of data that he may not look at or she may not look at from that particular. Particular same example, same use case, or a similar one, and so the idea is to be helpful in better processing of the Kappas, as well the results in kappa, but the investigation and the root cause analysis, in order to really have true effectiveness and corrective action that’s that’s demonstrating now in that particular instance, but when compared over time and and then through the use of AI tools. Can give you a future state of performance. I think we’re out of time. It took longer than expected, and sorry they moved some of the speech times around from yesterday to today. I wasn’t sure what I was going to be presenting, but thank you for your time. Enjoy your evening. Get to check out Raleigh, and then enjoy your drink if you do. Thank you.

← O’Brien on intended use | Home | End →

Graph View

Backlinks