Doc AI Transforms Cytokine Study Data

Amara Lawson

Hey y’all, welcome back to Deep Dive 360. I’m Amara, and I’m here with Ravi. Today, we’re talking about something that honestly gives me flashbacks—data chaos in cytokine research. Ravi, you ever open a folder and just see, like, three hundred Excel files staring back at you?

Ravi Kumar

Oh, absolutely. And not just three hundred files, but three hundred files with no rhyme or reason. I mean, we had a client prepping for a funding round, and their CRO sent them all these Excel sheets—every single one formatted differently. Some had merged cells, some had columns with names that made no sense, and sometimes you’d find two or three studies crammed into one sheet. It was a mess.

Amara Lawson

That’s the kind of thing that’ll make you want to just close your laptop and walk away. And it’s not just annoying, right? It’s risky. If you’re trying to get funding or hit a submission deadline, you can’t afford to spend weeks just cleaning up spreadsheets. I remember this one supplier audit—totally different industry, but same vibe. We had mismatched spreadsheets from three different plants, and it stalled us for weeks. Folks were just emailing back and forth, trying to figure out which version was the real one. It’s wild how universal this problem is.

Ravi Kumar

Exactly. And in research, that delay isn’t just a headache—it can mean missing out on funding, or not having the right data for a regulatory submission. Manual data management just slows everything down. You’ve got analysts spending weeks, sometimes months, just trying to get the data into a usable format. And that’s assuming there are no mistakes, which, let’s be honest, there always are.

Amara Lawson

Yeah, and the more hands in the pot, the more likely something’s gonna get missed or mixed up. It’s like playing telephone, but with numbers and deadlines.

Ravi Kumar

So, here’s where AI comes in. We worked with this client using our Doc AI Data Analytics module. The first thing we did was train the model to actually understand the context of each column—even when the columns weren’t named the same way. We had to annotate a bunch of samples, teach it to spot patterns in merged cells, and deal with all those non-uniform tables. Out of the box, we got about seventy percent of the data extracted automatically.

Amara Lawson

Wait, so what about the other thirty percent? That’s always the tricky part, right?

Ravi Kumar

Yeah, that’s where it gets interesting. Those files had no consistent structure—no anchors for the AI to grab onto. So, we brought in the company’s internal study protocol documents, matched them up with experiment numbers, and gave the model some clinical context. Suddenly, it started to make sense of the rest. We could extract the remaining data with a lot more confidence.

Amara Lawson

That’s pretty slick. But, I mean, this is critical stuff. How do you make sure the AI isn’t just making things up or missing something important?

Ravi Kumar

That’s where the human-in-the-loop comes in. We set up a quality control step using an AQL-based statistical sampling plan. Basically, we’d verify sample batches, audit the extracted records against the raw inputs, and make sure everything was compliance-grade before finalizing anything. It’s not just a black box spitting out numbers. There’s transparency, audit trails, and the client can review every step.

Amara Lawson

So, it’s not just “trust the robot”—it’s more like, “let the robot do the heavy lifting, but keep a human eye on the details.” I like that. Especially with sensitive data, you can’t just automate and hope for the best.

Ravi Kumar

Exactly. And privacy is huge in biosciences. We ran the whole pipeline in a secure, client-approved environment—no data leakage, no surprises. It’s all about balancing the speed and power of automation with the trust and oversight only humans can provide.

Amara Lawson

So, once you’ve got all this structured data, what’s next? I mean, it’s great to have clean spreadsheets, but folks want answers, not just files, right?

Ravi Kumar

Right. That’s where we took it a step further. We connected the structured data into our Doc AI RAG model—retrieval-augmented generation. Now, the client’s team can literally chat with their cytokine data. They can ask questions like, “Which studies had IL-6 levels above baseline at 48 hours?” or generate plots and comparisons instantly. No more digging through thousands of rows by hand.

Amara Lawson

That’s a game-changer. I mean, I’ve spent hours—okay, days—just trying to wrangle data into something I could actually use. Being able to just ask a question and get a plot back? That’s next-level.

Ravi Kumar

And the best part is, it’s not just a one-off. We’re building custom connectors for their internal drives, so future uploads go straight to structured outputs. And we’re seeing this work with other biotech and CRO clients, too. Different file formats, different study logic—Doc AI adapts by training on their data. It’s flexible.

Amara Lawson

So, what’s next? I mean, if AI can handle this kind of messy, high-stakes data, where do you see it going for data management in general?

Ravi Kumar

Honestly, I think we’re just scratching the surface. As more teams trust AI to handle the grunt work, they’ll be able to focus on the science, not the spreadsheets. And as the models get better at understanding context, we’ll see even more interactive, real-time analysis—maybe even predictive insights down the line. But, you know, always with that human oversight. That’s not going away.

Amara Lawson

Yeah, I’m with you. It’s about making the tech work for people, not the other way around. Well, that’s all we’ve got for today, folks. Ravi, thanks for breaking it down with me.

Ravi Kumar

Always