GxP compliance and data science career pathways discussion
Recorded 2026-09-16 at 1237 via Plaud. 18:06.
Summary
A hallway conversation in two halves. First, GxP enforcement stories from a manufacturing-side practitioner: employees fired for backdating documentation by a single day, FDA investigators collecting trash cans and finding sticky-note calculations from technicians who were forbidden to carry writing instruments, an FDA threat to withhold future drug approvals over repeated minor infractions, and an Indian generic-drug manufacturer driven to closure by a consent decree, after which its US multinational customer hired lawyers and roughly fifty private consultants to re-review twenty years of clinical and validation packages over data integrity. Eric offered his framing that GxP exists to build institutional trust on behalf of patients, and said the documentation density of GxP digital platforms finally made sense to him.
The second half is Eric’s career arc: undergraduate training in microbiology, biochemistry, and synthetic biology; self-taught computing during a PhD on influenza (Python in 2013); Novartis Scientific Data Analysis from 2017 to 2021, including a 2019 system that fitted 14,000 machine learning models in one run across the chemical activity database; deep-learning fundamentals learned sitting in on David Duvenaud’s Harvard lab on Friday afternoons; and the learning retreat he is organizing during his employer’s one-month sabbatical.
Plaud AI notes
<App-generated notes from the Plaud cloud, verbatim.>
This conversation covers regulatory compliance and data integrity within GXP and FDA-regulated environments, followed by an overview of Eric Ma’s transition from biological sciences into data science and machine learning. Speaker 1 illustrates how minor procedural violations can trigger severe regulatory penalties, while Eric Ma details his professional background, self-directed learning approach, and plans for an AI-focused learning retreat.
GXP Regulatory Compliance and FDA Data Integrity Enforcement
FDA regulatory oversight relies strictly on institutional trust, where minor deviations in documentation or procedure can jeopardize drug approvals.
- Documentation backdating: Employees were terminated for performing work on one day and signing the required GXP documentation the previous day.
- Unapproved record-keeping: During facility inspections, FDA investigators searched trash cans and found sticky notes containing manual calculations made by manufacturing technicians who were prohibited from carrying writing materials.
- Regulatory consequences: The FDA threatened to withhold future approvals for multi-billion-dollar drug pipelines due to recurring documentation infractions.
- Legal consent decrees: Following court action against an Indian generic drug manufacturer that led to company closure, a US multinational pharma had to hire lawyers and roughly 50 private consultants to review 20 years of clinical trial and validation packages for data integrity.
Even if the FDA cannot audit every record, discovery of a single minor infraction can prompt extensive institutional reviews. Eric Ma noted that these strict regulatory requirements explain why digital GXP platforms require explicit, granular documentation to ensure defensibility and patient safety.
Career Transition and Machine Learning Development
Eric Ma transitioned from an undergraduate background in microbiology, biochemistry, and synthetic biology into computational work during graduate school. He learned Python in 2013 while writing his PhD thesis on influenza, developing expertise in large-scale computing, statistical methods, machine learning, and deep learning through practical projects and postdoc mentorship rather than formal coursework.
- Novartis (2017–2021): Joined the Scientific Data Analysis team in 2017 after completing graduate school, where he built a system in 2019 that fitted 14,000 machine learning models in one run across the chemical activity database.
- Deep learning mentorship: Learned core concepts like backpropagation through informal mentorship sessions with David Duvenaud, then a Harvard postdoc who later became a professor at the University of Toronto.
- Statistical influences: Tracked prominent quantitative researchers, including Stanford statistician Rob Tibshirani and machine learning faculty at Cambridge University.
Professional Networking and AI Learning Retreat Initiative
Eric Ma is organizing an AI learning retreat planned for the following year during Moderna’s one-month employee sabbatical. The retreat aims to teach participants how to direct their own learning using AI and apply practical prompting techniques to convert AI outputs into thought-provoking tools.
During the discussion, Speaker 1 asked about Moderna researcher Juliana Jung, whom Eric Ma located and confirmed in his contact network. Both participants exchanged contact details before concluding the conversation as a separate group presentation began.
Transcript
[00:00] Speaker 1: Really? Didn’t know why they got fired. Was such a like minor thing. Yeah. Like they just backdated. Oh. I mean like they did um they didn’t do the work. The the next day they they signed it the day before. So as long as I thought that like as long as they remember like what they did and then like and signed one day later, that’s a problem. Right. And then like it was not like. And then when GMP, when FDA come here for approval of drug, and we have basically, they have like they’re filling up like, so it’s we have the trust. So it kind of like if you violate the trust, if you violate this, what would you violate? Elsewhere, because same principle applies. This event. Especially AI part, it’s kind of have to have to under trust job that the agencies have to show everything. Yeah,
[01:10] Eric Ma: Is it correct to say that the whole premise of GXB, because it’s about patient safety, therefore FDA needs to trust. We need to build trust with FDA, who’s acting on behalf behalf of the patients that we’re doing according to. Procedures that we’ve established that we believe are defensible.
[01:29] Speaker 1: Everything about Skype.
[01:31] Eric Ma: Okay. Okay. I under. I think I am understanding it better now, because I also like when I first took a look at like all the documentation that was needed for a GXP plat digital platform. I was actually very confused why it needs to be so so heavily documented. Why there’s so much emphasis on, All the nitty - gritty, minut, minut, minutiae, and and I think I’m starting to have a better appreciation for that now.
[02:03] Speaker 1: For me, it’s probably a lot more manufacturing, as you see, FDA, visitor us. They just shake up our trash can. They collect everything in the trash can.
[02:20] Eric Ma: As well, really.
[02:23] Speaker 1: And then they. In the trash can, you know, there’s a there’s a little little like sticky. Someone wrote the things on the sticky and throw away. They found out that you know this calculation is done in this sticky, and then they found that like they don’t want the manufacturing technicians. They write down. Because there’s, they don’t allow to bring riding thing. They cannot ride anywhere. Yeah. They don’t want to forget that they use their, you know, their gloves. Gloves. Right. It’s all, you know, that looks minor. You. Know, it. Almost 100% degree that they said if you keep doing this, we’re not gonna give you any new drug approval. I mean, like this this the current drug approval, and then like you were saying that you know multi-billion-dollar new drugs on on the pipeline. FDA saying, uh-oh, not for this, but for future drug, and that’s why that’s why. Okay. Yeah.
[03:46] Eric Ma: Definitely, it feels like there’s a big culture difference. That’s for sure, compared to research, which is more free wheeling.
[03:59] Speaker 1: Yeah, and also, not my company. Tenentech is really good and helps about GMP kind of because they they have resources, but they are. I have my consternacy. I can’t say which company, but they all this. It’s Indian company related in Indian manufacturers. And several several years, they brought this company into the court. FDA can brought. The consent decree is about the legal. Yeah. They can drop this company into court, so the company closed. Oh really? Company closed. There’s like thousands of like there’s a thousand yeah thousands. It’s quite big. And then the U S multinational company who hired this Indian company. They have to depend on twenty years. Twenty years. So imagine you have to depend on twenty years of clinical. So clinical. So this, this is this company is about the manufacturing of. No, no, it’s it’s more for. Generic drug. The basically, this generic drug has the clinical trial package, and all the documents, all the equivalence packing package. So they show all the clinical trials. So each each clinical trial. Of course, it’s not the size of original drug, but still the clinical trial, all the validation, all the report, 20 years. So how many drops, and then how much money, and then so this multinational pharma hired these lawyers, and then also like all our like a. Five zero different, like many many private, like me, private consultants to review review entire twenty years again. Wow. Imagine that with data integrity. Yeah. Single word. Yeah. Data integrity.
[06:53] Eric Ma: Data integrity.
[06:54] Speaker 1: Okay. Yeah.
[06:55] Eric Ma: Wow. It’s definitely like, definitely like a big change. That’s definitely a big change. The amount of meat, the consequences are very different as.
[07:20] Speaker 1: Well. You don’t, of course, you don’t get to have. I mean, even though you don’t have a big, poor. Big things. FDA cannot scan everything, but if if they pick one minor, they can kind of respond. Now.
[07:41] Eric Ma: I heard that as well.
[07:46] Speaker 1: About yes, how do you get into a computer computer background? Oh. I was in.
[07:59] Eric Ma: I was trained in undergrad microbiology, biochemistry, synthetic biology. I went to grad school through the back door of a lab that moved into the institute, and then when I was done, I had already switched advisors, and I did. I picked up computing on my own. In grad school, there’s there’s plenty of time during a PhD to pick up computing. Maybe not. I don’t know. Like lucky enough, lucky enough to have the space and time.
[08:40] Speaker 1: I think it’s microbiology. I was not able to in the transition that you did, but now I did.
[08:50] Eric Ma: Oh, that’s awesome! Yeah, now now I feel like it’s a lot easier to. Pick up a new topic with AI help like that’s much, much, much easier. There’s actually a like for side sidebar like there’s a retreat that I’m organizing next year during our Moderna has sabbatical months that they give to every employee, and it’s one month of just like we can do whatever we want. Oh yeah. So I discussed with my wife and I told her I want to do this retreat where we bring people who are who want to learn anything, who want to learn how to learn. Learn, learn how to learn with AI. That’s the retreat that I’m organizing, and so then like it’s just like very curious people who are curious and autodidactic. They can teach themselves stuff, but they want to beefen up their ability to learn, and then also teach the prompting, practical prompting, the methods for flipping AI from a slop generator to a thought provoker. I want to want to bring this group together. Yeah. Yeah.
[09:55] Speaker 1: Is it okay if I introduce you? I mean, by kind of say that my friend works in Madonna Research. She’s in research. Oh, what’s her name? Juliana Jung. Juliana Jung. I think Madonna. We’re small enough. I think she’s she’s she’s Juliana. I’m not so sure that she used Juliana. Oh,
[10:19] Eric Ma: She’s here. Let me just check to make sure I have. Is she still around? Let me just double check. People. She is. This must be her. Juliana Jones. That’s right. She put something. Yeah. The anyways, I I’m I’m organizing that retreat. I want to like bring people together. Want to learn anything? Yeah. Happy to. Happy to. If you want to connect on LinkedIn, I’m also happy to as well. Is there an iPhone? Yeah. Yeah. Let’s see if the bump works. The sound is. I heard bumping might work with it. Okay. I can share. I share. Share my contact. Or fair. Okay. Yeah. All right. Nice to meet you, Barry. Thanks a lot. Young, young, young man. Yeah. So, anyways, I got into data science from biology. I kind of like picked up Python in 2013. And it was just the perfect time. So I needed to write a thesis. I came up with the thesis, got a lot of help from other committee members on my committee, and then that’s roughly when I saw that this thing called data science was exploding. And so I was like, hey, let’s give this a shot. So I formulated my thesis mostly around like influenza, but also. Getting a lot of practice with large-scale computing and statistical methods, and machine learning and deep learning and that sort of stuff. So I trained myself over the years. You trained yourself? Mostly, I had I had help. Like I didn’t, I couldn’t take courses. Like I I don’t suit a classroom environment very well. Classroom can sometimes help if it’s very hands-on, but like if it’s mostly theory and just like it goes one year in, one year out. But I had I had I’ve kind of figured out how to make up projects for myself that could really test my understanding, and and that was helpful. And I had help from like postdocs and other people who were more experienced than myself as well. So that was cool. It was cool to to pick that up. And then I joined Novartis after I finished grad school. Twenty seventeen, I finished, and then. I was in a team called Scientific Data Analysis. It was like a SWAT team. We go in and just like help wherever, complement others, but we came in with specialized knowledge as well. And so we do. I did a lot of, lot of different projects at Novartis. One was on a system that fit fourteen thousand machine learning models at one shot across our entire chemical activity database. It it took a bit back then to create, but I think now if I to recreate it, no, twenty nineteen, twenty twenty nineteen, yeah, twenty nineteen, yeah, that’s right, before AI.
[14:10] Speaker 1: AI, presentation, studied like when I was in Stanford, the year two thousand, Rob Tibsirani, Stanford
[14:22] Eric Ma: Khatranee, yes. I heard of him. I’ve heard of him. I don’t know him, but I’ve heard of him.
[14:29] Speaker 1: He comes to our nap. Big hit up too. This is our. Every time we have a lot of Jin, micro range Jin. You know the Pat Patrick Brown. That one’s on.
[14:51] Eric Ma: Yeah. Yeah. Yeah, he’s he’s still working as a statistician. Yeah, yeah. Where where is he? Is still at Stanford, right? Yeah. Okay. Okay. Because I remember there’s like there’s a core of quantitative people I was paying attention to. Rob Tibshirani was one, and then in the Cambridge University when I was in grad school, Cambridge University had like two professors in machine learning that was like just like they were deities, you know? And they were and they were like big heroes. Yeah. And I was like paying attention to their work. It was it was pretty fun, to it was pretty fun to follow along. Yeah.
[15:30] Speaker 1: Really, I’m so impressed.
[15:32] Eric Ma: Thank you. I had I had good help. Like there were good good people in my life to to help with the coaching. Like for deep learning, for example, there was a postdoc at Harvard. He was in a lab of I forgot what the lab name, but the the postdoc name is David Durvanod, and, He turned out to be a pretty rising star in the in deep learning, and when he was a postdoc, I was a grad student, and I would just go ask him questions in my final year of grad school. I was like, David, I don’t understand this, right? Can you explain backprop to me? Like, I don’t fully understand it. And I was like, I I would like spend Friday afternoons sitting at at his lab. You know, I just bike up, forget my lab, I go to his lab, his his lab group, and sit there. Sit there and learn, learn from him. So then he taught me a lot of fundamentals from scratch. So I consider him to be like my my deep learning teacher. Yeah, it was it was great. David Duvenaud was like so good. Then he became a professor at U of T, and then had a few really cool papers at NeurIPS and so on. Yeah, yeah, yeah. Duvenaud is D U V E N A U D. David Duvanad. He was great. I still remember him teaching me a lot about backdrops. D u v e here I can type.
[17:17] Speaker 1: Sorry, it’s background noise. It’s okay. There’s going to be a presentation.
[17:24] Eric Ma: Oh, okay. Before he made, before he grew his hair and beard. I see. I see. Yeah.
[17:32] Speaker 1: So I am interested in his appliance group presenting their. I think I really like all the work I spent. Month. They just made it.
[17:51] Eric Ma: Check it out. Definitely. I think they’re probably inside, if I’m not mistaken. It’s probably on the inside. That’s good to meet you, Young Ran.